1 // SPDX-License-Identifier: GPL-2.0 1 // SPDX-License-Identifier: GPL-2.0
2 /* 2 /*
3 * Deadline Scheduling Class (SCHED_DEADLINE) 3 * Deadline Scheduling Class (SCHED_DEADLINE)
4 * 4 *
5 * Earliest Deadline First (EDF) + Constant Ba 5 * Earliest Deadline First (EDF) + Constant Bandwidth Server (CBS).
6 * 6 *
7 * Tasks that periodically executes their inst 7 * Tasks that periodically executes their instances for less than their
8 * runtime won't miss any of their deadlines. 8 * runtime won't miss any of their deadlines.
9 * Tasks that are not periodic or sporadic or 9 * Tasks that are not periodic or sporadic or that tries to execute more
10 * than their reserved bandwidth will be slowe 10 * than their reserved bandwidth will be slowed down (and may potentially
11 * miss some of their deadlines), and won't af 11 * miss some of their deadlines), and won't affect any other task.
12 * 12 *
13 * Copyright (C) 2012 Dario Faggioli <raistlin 13 * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
14 * Juri Lelli <juri.lelli@g 14 * Juri Lelli <juri.lelli@gmail.com>,
15 * Michael Trimarchi <micha 15 * Michael Trimarchi <michael@amarulasolutions.com>,
16 * Fabio Checconi <fcheccon 16 * Fabio Checconi <fchecconi@gmail.com>
17 */ 17 */
>> 18 #include "sched.h"
>> 19 #include "pelt.h"
18 20
19 #include <linux/cpuset.h> !! 21 struct dl_bandwidth def_dl_bandwidth;
20 <<
21 /* <<
22 * Default limits for DL period; on the top en <<
23 * tasks still getting ridiculously long effec <<
24 * guard against timer DoS. <<
25 */ <<
26 static unsigned int sysctl_sched_dl_period_max <<
27 static unsigned int sysctl_sched_dl_period_min <<
28 #ifdef CONFIG_SYSCTL <<
29 static struct ctl_table sched_dl_sysctls[] = { <<
30 { <<
31 .procname = "sched_deadl <<
32 .data = &sysctl_sche <<
33 .maxlen = sizeof(unsig <<
34 .mode = 0644, <<
35 .proc_handler = proc_douintv <<
36 .extra1 = (void *)&sys <<
37 }, <<
38 { <<
39 .procname = "sched_deadl <<
40 .data = &sysctl_sche <<
41 .maxlen = sizeof(unsig <<
42 .mode = 0644, <<
43 .proc_handler = proc_douintv <<
44 .extra2 = (void *)&sys <<
45 }, <<
46 }; <<
47 <<
48 static int __init sched_dl_sysctl_init(void) <<
49 { <<
50 register_sysctl_init("kernel", sched_d <<
51 return 0; <<
52 } <<
53 late_initcall(sched_dl_sysctl_init); <<
54 #endif <<
55 <<
56 static bool dl_server(struct sched_dl_entity * <<
57 { <<
58 return dl_se->dl_server; <<
59 } <<
60 22
61 static inline struct task_struct *dl_task_of(s 23 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
62 { 24 {
63 BUG_ON(dl_server(dl_se)); <<
64 return container_of(dl_se, struct task 25 return container_of(dl_se, struct task_struct, dl);
65 } 26 }
66 27
67 static inline struct rq *rq_of_dl_rq(struct dl 28 static inline struct rq *rq_of_dl_rq(struct dl_rq *dl_rq)
68 { 29 {
69 return container_of(dl_rq, struct rq, 30 return container_of(dl_rq, struct rq, dl);
70 } 31 }
71 32
72 static inline struct rq *rq_of_dl_se(struct sc <<
73 { <<
74 struct rq *rq = dl_se->rq; <<
75 <<
76 if (!dl_server(dl_se)) <<
77 rq = task_rq(dl_task_of(dl_se) <<
78 <<
79 return rq; <<
80 } <<
81 <<
82 static inline struct dl_rq *dl_rq_of_se(struct 33 static inline struct dl_rq *dl_rq_of_se(struct sched_dl_entity *dl_se)
83 { 34 {
84 return &rq_of_dl_se(dl_se)->dl; !! 35 struct task_struct *p = dl_task_of(dl_se);
>> 36 struct rq *rq = task_rq(p);
>> 37
>> 38 return &rq->dl;
85 } 39 }
86 40
87 static inline int on_dl_rq(struct sched_dl_ent 41 static inline int on_dl_rq(struct sched_dl_entity *dl_se)
88 { 42 {
89 return !RB_EMPTY_NODE(&dl_se->rb_node) 43 return !RB_EMPTY_NODE(&dl_se->rb_node);
90 } 44 }
91 45
92 #ifdef CONFIG_RT_MUTEXES <<
93 static inline struct sched_dl_entity *pi_of(st <<
94 { <<
95 return dl_se->pi_se; <<
96 } <<
97 <<
98 static inline bool is_dl_boosted(struct sched_ <<
99 { <<
100 return pi_of(dl_se) != dl_se; <<
101 } <<
102 #else <<
103 static inline struct sched_dl_entity *pi_of(st <<
104 { <<
105 return dl_se; <<
106 } <<
107 <<
108 static inline bool is_dl_boosted(struct sched_ <<
109 { <<
110 return false; <<
111 } <<
112 #endif <<
113 <<
114 #ifdef CONFIG_SMP 46 #ifdef CONFIG_SMP
115 static inline struct dl_bw *dl_bw_of(int i) 47 static inline struct dl_bw *dl_bw_of(int i)
116 { 48 {
117 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ 49 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
118 "sched RCU must be he 50 "sched RCU must be held");
119 return &cpu_rq(i)->rd->dl_bw; 51 return &cpu_rq(i)->rd->dl_bw;
120 } 52 }
121 53
122 static inline int dl_bw_cpus(int i) 54 static inline int dl_bw_cpus(int i)
123 { 55 {
124 struct root_domain *rd = cpu_rq(i)->rd 56 struct root_domain *rd = cpu_rq(i)->rd;
125 int cpus; 57 int cpus;
126 58
127 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ 59 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
128 "sched RCU must be he 60 "sched RCU must be held");
129 61
130 if (cpumask_subset(rd->span, cpu_activ 62 if (cpumask_subset(rd->span, cpu_active_mask))
131 return cpumask_weight(rd->span 63 return cpumask_weight(rd->span);
132 64
133 cpus = 0; 65 cpus = 0;
134 66
135 for_each_cpu_and(i, rd->span, cpu_acti 67 for_each_cpu_and(i, rd->span, cpu_active_mask)
136 cpus++; 68 cpus++;
137 69
138 return cpus; 70 return cpus;
139 } 71 }
140 72
141 static inline unsigned long __dl_bw_capacity(c !! 73 static inline unsigned long __dl_bw_capacity(int i)
142 { 74 {
>> 75 struct root_domain *rd = cpu_rq(i)->rd;
143 unsigned long cap = 0; 76 unsigned long cap = 0;
144 int i; <<
145 77
146 for_each_cpu_and(i, mask, cpu_active_m !! 78 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
147 cap += arch_scale_cpu_capacity !! 79 "sched RCU must be held");
>> 80
>> 81 for_each_cpu_and(i, rd->span, cpu_active_mask)
>> 82 cap += capacity_orig_of(i);
148 83
149 return cap; 84 return cap;
150 } 85 }
151 86
152 /* 87 /*
153 * XXX Fix: If 'rq->rd == def_root_domain' per 88 * XXX Fix: If 'rq->rd == def_root_domain' perform AC against capacity
154 * of the CPU the task is running on rather rd 89 * of the CPU the task is running on rather rd's \Sum CPU capacity.
155 */ 90 */
156 static inline unsigned long dl_bw_capacity(int 91 static inline unsigned long dl_bw_capacity(int i)
157 { 92 {
158 if (!sched_asym_cpucap_active() && !! 93 if (!static_branch_unlikely(&sched_asym_cpucapacity) &&
159 arch_scale_cpu_capacity(i) == SCHE !! 94 capacity_orig_of(i) == SCHED_CAPACITY_SCALE) {
160 return dl_bw_cpus(i) << SCHED_ 95 return dl_bw_cpus(i) << SCHED_CAPACITY_SHIFT;
161 } else { 96 } else {
162 RCU_LOCKDEP_WARN(!rcu_read_loc !! 97 return __dl_bw_capacity(i);
163 "sched RCU mu <<
164 <<
165 return __dl_bw_capacity(cpu_rq <<
166 } <<
167 } <<
168 <<
169 static inline bool dl_bw_visited(int cpu, u64 <<
170 { <<
171 struct root_domain *rd = cpu_rq(cpu)-> <<
172 <<
173 if (rd->visit_gen == gen) <<
174 return true; <<
175 <<
176 rd->visit_gen = gen; <<
177 return false; <<
178 } <<
179 <<
180 static inline <<
181 void __dl_update(struct dl_bw *dl_b, s64 bw) <<
182 { <<
183 struct root_domain *rd = container_of( <<
184 int i; <<
185 <<
186 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_ <<
187 "sched RCU must be he <<
188 for_each_cpu_and(i, rd->span, cpu_acti <<
189 struct rq *rq = cpu_rq(i); <<
190 <<
191 rq->dl.extra_bw += bw; <<
192 } 98 }
193 } 99 }
194 #else 100 #else
195 static inline struct dl_bw *dl_bw_of(int i) 101 static inline struct dl_bw *dl_bw_of(int i)
196 { 102 {
197 return &cpu_rq(i)->dl.dl_bw; 103 return &cpu_rq(i)->dl.dl_bw;
198 } 104 }
199 105
200 static inline int dl_bw_cpus(int i) 106 static inline int dl_bw_cpus(int i)
201 { 107 {
202 return 1; 108 return 1;
203 } 109 }
204 110
205 static inline unsigned long dl_bw_capacity(int 111 static inline unsigned long dl_bw_capacity(int i)
206 { 112 {
207 return SCHED_CAPACITY_SCALE; 113 return SCHED_CAPACITY_SCALE;
208 } 114 }
209 <<
210 static inline bool dl_bw_visited(int cpu, u64 <<
211 { <<
212 return false; <<
213 } <<
214 <<
215 static inline <<
216 void __dl_update(struct dl_bw *dl_b, s64 bw) <<
217 { <<
218 struct dl_rq *dl = container_of(dl_b, <<
219 <<
220 dl->extra_bw += bw; <<
221 } <<
222 #endif 115 #endif
223 116
224 static inline 117 static inline
225 void __dl_sub(struct dl_bw *dl_b, u64 tsk_bw, <<
226 { <<
227 dl_b->total_bw -= tsk_bw; <<
228 __dl_update(dl_b, (s32)tsk_bw / cpus); <<
229 } <<
230 <<
231 static inline <<
232 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw, <<
233 { <<
234 dl_b->total_bw += tsk_bw; <<
235 __dl_update(dl_b, -((s32)tsk_bw / cpus <<
236 } <<
237 <<
238 static inline bool <<
239 __dl_overflow(struct dl_bw *dl_b, unsigned lon <<
240 { <<
241 return dl_b->bw != -1 && <<
242 cap_scale(dl_b->bw, cap) < dl_b <<
243 } <<
244 <<
245 static inline <<
246 void __add_running_bw(u64 dl_bw, struct dl_rq 118 void __add_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
247 { 119 {
248 u64 old = dl_rq->running_bw; 120 u64 old = dl_rq->running_bw;
249 121
250 lockdep_assert_rq_held(rq_of_dl_rq(dl_ !! 122 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
251 dl_rq->running_bw += dl_bw; 123 dl_rq->running_bw += dl_bw;
252 SCHED_WARN_ON(dl_rq->running_bw < old) 124 SCHED_WARN_ON(dl_rq->running_bw < old); /* overflow */
253 SCHED_WARN_ON(dl_rq->running_bw > dl_r 125 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
254 /* kick cpufreq (see the comment in ke 126 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
255 cpufreq_update_util(rq_of_dl_rq(dl_rq) 127 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
256 } 128 }
257 129
258 static inline 130 static inline
259 void __sub_running_bw(u64 dl_bw, struct dl_rq 131 void __sub_running_bw(u64 dl_bw, struct dl_rq *dl_rq)
260 { 132 {
261 u64 old = dl_rq->running_bw; 133 u64 old = dl_rq->running_bw;
262 134
263 lockdep_assert_rq_held(rq_of_dl_rq(dl_ !! 135 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
264 dl_rq->running_bw -= dl_bw; 136 dl_rq->running_bw -= dl_bw;
265 SCHED_WARN_ON(dl_rq->running_bw > old) 137 SCHED_WARN_ON(dl_rq->running_bw > old); /* underflow */
266 if (dl_rq->running_bw > old) 138 if (dl_rq->running_bw > old)
267 dl_rq->running_bw = 0; 139 dl_rq->running_bw = 0;
268 /* kick cpufreq (see the comment in ke 140 /* kick cpufreq (see the comment in kernel/sched/sched.h). */
269 cpufreq_update_util(rq_of_dl_rq(dl_rq) 141 cpufreq_update_util(rq_of_dl_rq(dl_rq), 0);
270 } 142 }
271 143
272 static inline 144 static inline
273 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_r 145 void __add_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
274 { 146 {
275 u64 old = dl_rq->this_bw; 147 u64 old = dl_rq->this_bw;
276 148
277 lockdep_assert_rq_held(rq_of_dl_rq(dl_ !! 149 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
278 dl_rq->this_bw += dl_bw; 150 dl_rq->this_bw += dl_bw;
279 SCHED_WARN_ON(dl_rq->this_bw < old); / 151 SCHED_WARN_ON(dl_rq->this_bw < old); /* overflow */
280 } 152 }
281 153
282 static inline 154 static inline
283 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_r 155 void __sub_rq_bw(u64 dl_bw, struct dl_rq *dl_rq)
284 { 156 {
285 u64 old = dl_rq->this_bw; 157 u64 old = dl_rq->this_bw;
286 158
287 lockdep_assert_rq_held(rq_of_dl_rq(dl_ !! 159 lockdep_assert_held(&(rq_of_dl_rq(dl_rq))->lock);
288 dl_rq->this_bw -= dl_bw; 160 dl_rq->this_bw -= dl_bw;
289 SCHED_WARN_ON(dl_rq->this_bw > old); / 161 SCHED_WARN_ON(dl_rq->this_bw > old); /* underflow */
290 if (dl_rq->this_bw > old) 162 if (dl_rq->this_bw > old)
291 dl_rq->this_bw = 0; 163 dl_rq->this_bw = 0;
292 SCHED_WARN_ON(dl_rq->running_bw > dl_r 164 SCHED_WARN_ON(dl_rq->running_bw > dl_rq->this_bw);
293 } 165 }
294 166
295 static inline 167 static inline
296 void add_rq_bw(struct sched_dl_entity *dl_se, 168 void add_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
297 { 169 {
298 if (!dl_entity_is_special(dl_se)) 170 if (!dl_entity_is_special(dl_se))
299 __add_rq_bw(dl_se->dl_bw, dl_r 171 __add_rq_bw(dl_se->dl_bw, dl_rq);
300 } 172 }
301 173
302 static inline 174 static inline
303 void sub_rq_bw(struct sched_dl_entity *dl_se, 175 void sub_rq_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
304 { 176 {
305 if (!dl_entity_is_special(dl_se)) 177 if (!dl_entity_is_special(dl_se))
306 __sub_rq_bw(dl_se->dl_bw, dl_r 178 __sub_rq_bw(dl_se->dl_bw, dl_rq);
307 } 179 }
308 180
309 static inline 181 static inline
310 void add_running_bw(struct sched_dl_entity *dl 182 void add_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
311 { 183 {
312 if (!dl_entity_is_special(dl_se)) 184 if (!dl_entity_is_special(dl_se))
313 __add_running_bw(dl_se->dl_bw, 185 __add_running_bw(dl_se->dl_bw, dl_rq);
314 } 186 }
315 187
316 static inline 188 static inline
317 void sub_running_bw(struct sched_dl_entity *dl 189 void sub_running_bw(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
318 { 190 {
319 if (!dl_entity_is_special(dl_se)) 191 if (!dl_entity_is_special(dl_se))
320 __sub_running_bw(dl_se->dl_bw, 192 __sub_running_bw(dl_se->dl_bw, dl_rq);
321 } 193 }
322 194
323 static void dl_rq_change_utilization(struct rq !! 195 static void dl_change_utilization(struct task_struct *p, u64 new_bw)
324 { 196 {
325 if (dl_se->dl_non_contending) { !! 197 struct rq *rq;
326 sub_running_bw(dl_se, &rq->dl) !! 198
327 dl_se->dl_non_contending = 0; !! 199 BUG_ON(p->dl.flags & SCHED_FLAG_SUGOV);
>> 200
>> 201 if (task_on_rq_queued(p))
>> 202 return;
328 203
>> 204 rq = task_rq(p);
>> 205 if (p->dl.dl_non_contending) {
>> 206 sub_running_bw(&p->dl, &rq->dl);
>> 207 p->dl.dl_non_contending = 0;
329 /* 208 /*
330 * If the timer handler is cur 209 * If the timer handler is currently running and the
331 * timer cannot be canceled, i !! 210 * timer cannot be cancelled, inactive_task_timer()
332 * will see that dl_not_conten 211 * will see that dl_not_contending is not set, and
333 * will not touch the rq's act 212 * will not touch the rq's active utilization,
334 * so we are still safe. 213 * so we are still safe.
335 */ 214 */
336 if (hrtimer_try_to_cancel(&dl_ !! 215 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
337 if (!dl_server(dl_se)) !! 216 put_task_struct(p);
338 put_task_struc <<
339 } <<
340 } 217 }
341 __sub_rq_bw(dl_se->dl_bw, &rq->dl); !! 218 __sub_rq_bw(p->dl.dl_bw, &rq->dl);
342 __add_rq_bw(new_bw, &rq->dl); 219 __add_rq_bw(new_bw, &rq->dl);
343 } 220 }
344 221
345 static void dl_change_utilization(struct task_ <<
346 { <<
347 WARN_ON_ONCE(p->dl.flags & SCHED_FLAG_ <<
348 <<
349 if (task_on_rq_queued(p)) <<
350 return; <<
351 <<
352 dl_rq_change_utilization(task_rq(p), & <<
353 } <<
354 <<
355 static void __dl_clear_params(struct sched_dl_ <<
356 <<
357 /* 222 /*
358 * The utilization of a task cannot be immedia 223 * The utilization of a task cannot be immediately removed from
359 * the rq active utilization (running_bw) when 224 * the rq active utilization (running_bw) when the task blocks.
360 * Instead, we have to wait for the so called 225 * Instead, we have to wait for the so called "0-lag time".
361 * 226 *
362 * If a task blocks before the "0-lag time", a 227 * If a task blocks before the "0-lag time", a timer (the inactive
363 * timer) is armed, and running_bw is decrease 228 * timer) is armed, and running_bw is decreased when the timer
364 * fires. 229 * fires.
365 * 230 *
366 * If the task wakes up again before the inact 231 * If the task wakes up again before the inactive timer fires,
367 * the timer is canceled, whereas if the task !! 232 * the timer is cancelled, whereas if the task wakes up after the
368 * inactive timer fired (and running_bw has be 233 * inactive timer fired (and running_bw has been decreased) the
369 * task's utilization has to be added to runni 234 * task's utilization has to be added to running_bw again.
370 * A flag in the deadline scheduling entity (d 235 * A flag in the deadline scheduling entity (dl_non_contending)
371 * is used to avoid race conditions between th 236 * is used to avoid race conditions between the inactive timer handler
372 * and task wakeups. 237 * and task wakeups.
373 * 238 *
374 * The following diagram shows how running_bw 239 * The following diagram shows how running_bw is updated. A task is
375 * "ACTIVE" when its utilization contributes t 240 * "ACTIVE" when its utilization contributes to running_bw; an
376 * "ACTIVE contending" task is in the TASK_RUN 241 * "ACTIVE contending" task is in the TASK_RUNNING state, while an
377 * "ACTIVE non contending" task is a blocked t 242 * "ACTIVE non contending" task is a blocked task for which the "0-lag time"
378 * has not passed yet. An "INACTIVE" task is a 243 * has not passed yet. An "INACTIVE" task is a task for which the "0-lag"
379 * time already passed, which does not contrib 244 * time already passed, which does not contribute to running_bw anymore.
380 * +------------- 245 * +------------------+
381 * wakeup | ACTIVE 246 * wakeup | ACTIVE |
382 * +------------------>+ contending 247 * +------------------>+ contending |
383 * | add_running_bw | 248 * | add_running_bw | |
384 * | +----+------+- 249 * | +----+------+------+
385 * | | ^ 250 * | | ^
386 * | dequeue | | 251 * | dequeue | |
387 * +--------+-------+ | | 252 * +--------+-------+ | |
388 * | | t >= 0-lag | | 253 * | | t >= 0-lag | | wakeup
389 * | INACTIVE |<---------------+ | 254 * | INACTIVE |<---------------+ |
390 * | | sub_running_bw | | 255 * | | sub_running_bw | |
391 * +--------+-------+ | | 256 * +--------+-------+ | |
392 * ^ | | 257 * ^ | |
393 * | t < 0-lag | | 258 * | t < 0-lag | |
394 * | | | 259 * | | |
395 * | V | 260 * | V |
396 * | +----+------+- 261 * | +----+------+------+
397 * | sub_running_bw | ACTIVE 262 * | sub_running_bw | ACTIVE |
398 * +-------------------+ 263 * +-------------------+ |
399 * inactive timer | non contend 264 * inactive timer | non contending |
400 * fired +------------- 265 * fired +------------------+
401 * 266 *
402 * The task_non_contending() function is invok 267 * The task_non_contending() function is invoked when a task
403 * blocks, and checks if the 0-lag time alread 268 * blocks, and checks if the 0-lag time already passed or
404 * not (in the first case, it directly updates 269 * not (in the first case, it directly updates running_bw;
405 * in the second case, it arms the inactive ti 270 * in the second case, it arms the inactive timer).
406 * 271 *
407 * The task_contending() function is invoked w 272 * The task_contending() function is invoked when a task wakes
408 * up, and checks if the task is still in the 273 * up, and checks if the task is still in the "ACTIVE non contending"
409 * state or not (in the second case, it update 274 * state or not (in the second case, it updates running_bw).
410 */ 275 */
411 static void task_non_contending(struct sched_d !! 276 static void task_non_contending(struct task_struct *p)
412 { 277 {
>> 278 struct sched_dl_entity *dl_se = &p->dl;
413 struct hrtimer *timer = &dl_se->inacti 279 struct hrtimer *timer = &dl_se->inactive_timer;
414 struct rq *rq = rq_of_dl_se(dl_se); !! 280 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
415 struct dl_rq *dl_rq = &rq->dl; !! 281 struct rq *rq = rq_of_dl_rq(dl_rq);
416 s64 zerolag_time; 282 s64 zerolag_time;
417 283
418 /* 284 /*
419 * If this is a non-deadline task that 285 * If this is a non-deadline task that has been boosted,
420 * do nothing 286 * do nothing
421 */ 287 */
422 if (dl_se->dl_runtime == 0) 288 if (dl_se->dl_runtime == 0)
423 return; 289 return;
424 290
425 if (dl_entity_is_special(dl_se)) 291 if (dl_entity_is_special(dl_se))
426 return; 292 return;
427 293
428 WARN_ON(dl_se->dl_non_contending); 294 WARN_ON(dl_se->dl_non_contending);
429 295
430 zerolag_time = dl_se->deadline - 296 zerolag_time = dl_se->deadline -
431 div64_long((dl_se->runtime * 297 div64_long((dl_se->runtime * dl_se->dl_period),
432 dl_se->dl_runtime); 298 dl_se->dl_runtime);
433 299
434 /* 300 /*
435 * Using relative times instead of the 301 * Using relative times instead of the absolute "0-lag time"
436 * allows to simplify the code 302 * allows to simplify the code
437 */ 303 */
438 zerolag_time -= rq_clock(rq); 304 zerolag_time -= rq_clock(rq);
439 305
440 /* 306 /*
441 * If the "0-lag time" already passed, 307 * If the "0-lag time" already passed, decrease the active
442 * utilization now, instead of startin 308 * utilization now, instead of starting a timer
443 */ 309 */
444 if ((zerolag_time < 0) || hrtimer_acti 310 if ((zerolag_time < 0) || hrtimer_active(&dl_se->inactive_timer)) {
445 if (dl_server(dl_se)) { !! 311 if (dl_task(p))
446 sub_running_bw(dl_se, 312 sub_running_bw(dl_se, dl_rq);
447 } else { !! 313 if (!dl_task(p) || p->state == TASK_DEAD) {
448 struct task_struct *p !! 314 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
449 315
450 if (dl_task(p)) !! 316 if (p->state == TASK_DEAD)
451 sub_running_bw !! 317 sub_rq_bw(&p->dl, &rq->dl);
452 !! 318 raw_spin_lock(&dl_b->lock);
453 if (!dl_task(p) || REA !! 319 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
454 struct dl_bw * !! 320 __dl_clear_params(p);
455 !! 321 raw_spin_unlock(&dl_b->lock);
456 if (READ_ONCE( <<
457 sub_rq <<
458 raw_spin_lock( <<
459 __dl_sub(dl_b, <<
460 raw_spin_unloc <<
461 __dl_clear_par <<
462 } <<
463 } 322 }
464 323
465 return; 324 return;
466 } 325 }
467 326
468 dl_se->dl_non_contending = 1; 327 dl_se->dl_non_contending = 1;
469 if (!dl_server(dl_se)) !! 328 get_task_struct(p);
470 get_task_struct(dl_task_of(dl_ <<
471 <<
472 hrtimer_start(timer, ns_to_ktime(zerol 329 hrtimer_start(timer, ns_to_ktime(zerolag_time), HRTIMER_MODE_REL_HARD);
473 } 330 }
474 331
475 static void task_contending(struct sched_dl_en 332 static void task_contending(struct sched_dl_entity *dl_se, int flags)
476 { 333 {
477 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 334 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
478 335
479 /* 336 /*
480 * If this is a non-deadline task that 337 * If this is a non-deadline task that has been boosted,
481 * do nothing 338 * do nothing
482 */ 339 */
483 if (dl_se->dl_runtime == 0) 340 if (dl_se->dl_runtime == 0)
484 return; 341 return;
485 342
486 if (flags & ENQUEUE_MIGRATED) 343 if (flags & ENQUEUE_MIGRATED)
487 add_rq_bw(dl_se, dl_rq); 344 add_rq_bw(dl_se, dl_rq);
488 345
489 if (dl_se->dl_non_contending) { 346 if (dl_se->dl_non_contending) {
490 dl_se->dl_non_contending = 0; 347 dl_se->dl_non_contending = 0;
491 /* 348 /*
492 * If the timer handler is cur 349 * If the timer handler is currently running and the
493 * timer cannot be canceled, i !! 350 * timer cannot be cancelled, inactive_task_timer()
494 * will see that dl_not_conten 351 * will see that dl_not_contending is not set, and
495 * will not touch the rq's act 352 * will not touch the rq's active utilization,
496 * so we are still safe. 353 * so we are still safe.
497 */ 354 */
498 if (hrtimer_try_to_cancel(&dl_ !! 355 if (hrtimer_try_to_cancel(&dl_se->inactive_timer) == 1)
499 if (!dl_server(dl_se)) !! 356 put_task_struct(dl_task_of(dl_se));
500 put_task_struc <<
501 } <<
502 } else { 357 } else {
503 /* 358 /*
504 * Since "dl_non_contending" i 359 * Since "dl_non_contending" is not set, the
505 * task's utilization has alre 360 * task's utilization has already been removed from
506 * active utilization (either 361 * active utilization (either when the task blocked,
507 * when the "inactive timer" f 362 * when the "inactive timer" fired).
508 * So, add it back. 363 * So, add it back.
509 */ 364 */
510 add_running_bw(dl_se, dl_rq); 365 add_running_bw(dl_se, dl_rq);
511 } 366 }
512 } 367 }
513 368
514 static inline int is_leftmost(struct sched_dl_ !! 369 static inline int is_leftmost(struct task_struct *p, struct dl_rq *dl_rq)
515 { 370 {
516 return rb_first_cached(&dl_rq->root) = !! 371 struct sched_dl_entity *dl_se = &p->dl;
>> 372
>> 373 return dl_rq->root.rb_leftmost == &dl_se->rb_node;
517 } 374 }
518 375
519 static void init_dl_rq_bw_ratio(struct dl_rq * 376 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq);
520 377
>> 378 void init_dl_bandwidth(struct dl_bandwidth *dl_b, u64 period, u64 runtime)
>> 379 {
>> 380 raw_spin_lock_init(&dl_b->dl_runtime_lock);
>> 381 dl_b->dl_period = period;
>> 382 dl_b->dl_runtime = runtime;
>> 383 }
>> 384
521 void init_dl_bw(struct dl_bw *dl_b) 385 void init_dl_bw(struct dl_bw *dl_b)
522 { 386 {
523 raw_spin_lock_init(&dl_b->lock); 387 raw_spin_lock_init(&dl_b->lock);
>> 388 raw_spin_lock(&def_dl_bandwidth.dl_runtime_lock);
524 if (global_rt_runtime() == RUNTIME_INF 389 if (global_rt_runtime() == RUNTIME_INF)
525 dl_b->bw = -1; 390 dl_b->bw = -1;
526 else 391 else
527 dl_b->bw = to_ratio(global_rt_ 392 dl_b->bw = to_ratio(global_rt_period(), global_rt_runtime());
>> 393 raw_spin_unlock(&def_dl_bandwidth.dl_runtime_lock);
528 dl_b->total_bw = 0; 394 dl_b->total_bw = 0;
529 } 395 }
530 396
531 void init_dl_rq(struct dl_rq *dl_rq) 397 void init_dl_rq(struct dl_rq *dl_rq)
532 { 398 {
533 dl_rq->root = RB_ROOT_CACHED; 399 dl_rq->root = RB_ROOT_CACHED;
534 400
535 #ifdef CONFIG_SMP 401 #ifdef CONFIG_SMP
536 /* zero means no -deadline tasks */ 402 /* zero means no -deadline tasks */
537 dl_rq->earliest_dl.curr = dl_rq->earli 403 dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
538 404
>> 405 dl_rq->dl_nr_migratory = 0;
539 dl_rq->overloaded = 0; 406 dl_rq->overloaded = 0;
540 dl_rq->pushable_dl_tasks_root = RB_ROO 407 dl_rq->pushable_dl_tasks_root = RB_ROOT_CACHED;
541 #else 408 #else
542 init_dl_bw(&dl_rq->dl_bw); 409 init_dl_bw(&dl_rq->dl_bw);
543 #endif 410 #endif
544 411
545 dl_rq->running_bw = 0; 412 dl_rq->running_bw = 0;
546 dl_rq->this_bw = 0; 413 dl_rq->this_bw = 0;
547 init_dl_rq_bw_ratio(dl_rq); 414 init_dl_rq_bw_ratio(dl_rq);
548 } 415 }
549 416
550 #ifdef CONFIG_SMP 417 #ifdef CONFIG_SMP
551 418
552 static inline int dl_overloaded(struct rq *rq) 419 static inline int dl_overloaded(struct rq *rq)
553 { 420 {
554 return atomic_read(&rq->rd->dlo_count) 421 return atomic_read(&rq->rd->dlo_count);
555 } 422 }
556 423
557 static inline void dl_set_overload(struct rq * 424 static inline void dl_set_overload(struct rq *rq)
558 { 425 {
559 if (!rq->online) 426 if (!rq->online)
560 return; 427 return;
561 428
562 cpumask_set_cpu(rq->cpu, rq->rd->dlo_m 429 cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
563 /* 430 /*
564 * Must be visible before the overload 431 * Must be visible before the overload count is
565 * set (as in sched_rt.c). 432 * set (as in sched_rt.c).
566 * 433 *
567 * Matched by the barrier in pull_dl_t 434 * Matched by the barrier in pull_dl_task().
568 */ 435 */
569 smp_wmb(); 436 smp_wmb();
570 atomic_inc(&rq->rd->dlo_count); 437 atomic_inc(&rq->rd->dlo_count);
571 } 438 }
572 439
573 static inline void dl_clear_overload(struct rq 440 static inline void dl_clear_overload(struct rq *rq)
574 { 441 {
575 if (!rq->online) 442 if (!rq->online)
576 return; 443 return;
577 444
578 atomic_dec(&rq->rd->dlo_count); 445 atomic_dec(&rq->rd->dlo_count);
579 cpumask_clear_cpu(rq->cpu, rq->rd->dlo 446 cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
580 } 447 }
581 448
582 #define __node_2_pdl(node) \ !! 449 static void update_dl_migration(struct dl_rq *dl_rq)
583 rb_entry((node), struct task_struct, p !! 450 {
>> 451 if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_running > 1) {
>> 452 if (!dl_rq->overloaded) {
>> 453 dl_set_overload(rq_of_dl_rq(dl_rq));
>> 454 dl_rq->overloaded = 1;
>> 455 }
>> 456 } else if (dl_rq->overloaded) {
>> 457 dl_clear_overload(rq_of_dl_rq(dl_rq));
>> 458 dl_rq->overloaded = 0;
>> 459 }
>> 460 }
584 461
585 static inline bool __pushable_less(struct rb_n !! 462 static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
586 { 463 {
587 return dl_entity_preempt(&__node_2_pdl !! 464 struct task_struct *p = dl_task_of(dl_se);
>> 465
>> 466 if (p->nr_cpus_allowed > 1)
>> 467 dl_rq->dl_nr_migratory++;
>> 468
>> 469 update_dl_migration(dl_rq);
588 } 470 }
589 471
590 static inline int has_pushable_dl_tasks(struct !! 472 static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
591 { 473 {
592 return !RB_EMPTY_ROOT(&rq->dl.pushable !! 474 struct task_struct *p = dl_task_of(dl_se);
>> 475
>> 476 if (p->nr_cpus_allowed > 1)
>> 477 dl_rq->dl_nr_migratory--;
>> 478
>> 479 update_dl_migration(dl_rq);
593 } 480 }
594 481
595 /* 482 /*
596 * The list of pushable -deadline task is not 483 * The list of pushable -deadline task is not a plist, like in
597 * sched_rt.c, it is an rb-tree with tasks ord 484 * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
598 */ 485 */
599 static void enqueue_pushable_dl_task(struct rq 486 static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
600 { 487 {
601 struct rb_node *leftmost; !! 488 struct dl_rq *dl_rq = &rq->dl;
602 !! 489 struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_root.rb_node;
603 WARN_ON_ONCE(!RB_EMPTY_NODE(&p->pushab !! 490 struct rb_node *parent = NULL;
>> 491 struct task_struct *entry;
>> 492 bool leftmost = true;
>> 493
>> 494 BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
>> 495
>> 496 while (*link) {
>> 497 parent = *link;
>> 498 entry = rb_entry(parent, struct task_struct,
>> 499 pushable_dl_tasks);
>> 500 if (dl_entity_preempt(&p->dl, &entry->dl))
>> 501 link = &parent->rb_left;
>> 502 else {
>> 503 link = &parent->rb_right;
>> 504 leftmost = false;
>> 505 }
>> 506 }
604 507
605 leftmost = rb_add_cached(&p->pushable_ <<
606 &rq->dl.pusha <<
607 __pushable_le <<
608 if (leftmost) 508 if (leftmost)
609 rq->dl.earliest_dl.next = p->d !! 509 dl_rq->earliest_dl.next = p->dl.deadline;
610 510
611 if (!rq->dl.overloaded) { !! 511 rb_link_node(&p->pushable_dl_tasks, parent, link);
612 dl_set_overload(rq); !! 512 rb_insert_color_cached(&p->pushable_dl_tasks,
613 rq->dl.overloaded = 1; !! 513 &dl_rq->pushable_dl_tasks_root, leftmost);
614 } <<
615 } 514 }
616 515
617 static void dequeue_pushable_dl_task(struct rq 516 static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
618 { 517 {
619 struct dl_rq *dl_rq = &rq->dl; 518 struct dl_rq *dl_rq = &rq->dl;
620 struct rb_root_cached *root = &dl_rq-> <<
621 struct rb_node *leftmost; <<
622 519
623 if (RB_EMPTY_NODE(&p->pushable_dl_task 520 if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
624 return; 521 return;
625 522
626 leftmost = rb_erase_cached(&p->pushabl !! 523 if (dl_rq->pushable_dl_tasks_root.rb_leftmost == &p->pushable_dl_tasks) {
627 if (leftmost) !! 524 struct rb_node *next_node;
628 dl_rq->earliest_dl.next = __no !! 525
>> 526 next_node = rb_next(&p->pushable_dl_tasks);
>> 527 if (next_node) {
>> 528 dl_rq->earliest_dl.next = rb_entry(next_node,
>> 529 struct task_struct, pushable_dl_tasks)->dl.deadline;
>> 530 }
>> 531 }
629 532
>> 533 rb_erase_cached(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
630 RB_CLEAR_NODE(&p->pushable_dl_tasks); 534 RB_CLEAR_NODE(&p->pushable_dl_tasks);
>> 535 }
631 536
632 if (!has_pushable_dl_tasks(rq) && rq-> !! 537 static inline int has_pushable_dl_tasks(struct rq *rq)
633 dl_clear_overload(rq); !! 538 {
634 rq->dl.overloaded = 0; !! 539 return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root.rb_root);
635 } <<
636 } 540 }
637 541
638 static int push_dl_task(struct rq *rq); 542 static int push_dl_task(struct rq *rq);
639 543
640 static inline bool need_pull_dl_task(struct rq 544 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
641 { 545 {
642 return rq->online && dl_task(prev); !! 546 return dl_task(prev);
643 } 547 }
644 548
645 static DEFINE_PER_CPU(struct balance_callback, !! 549 static DEFINE_PER_CPU(struct callback_head, dl_push_head);
646 static DEFINE_PER_CPU(struct balance_callback, !! 550 static DEFINE_PER_CPU(struct callback_head, dl_pull_head);
647 551
648 static void push_dl_tasks(struct rq *); 552 static void push_dl_tasks(struct rq *);
649 static void pull_dl_task(struct rq *); 553 static void pull_dl_task(struct rq *);
650 554
651 static inline void deadline_queue_push_tasks(s 555 static inline void deadline_queue_push_tasks(struct rq *rq)
652 { 556 {
653 if (!has_pushable_dl_tasks(rq)) 557 if (!has_pushable_dl_tasks(rq))
654 return; 558 return;
655 559
656 queue_balance_callback(rq, &per_cpu(dl 560 queue_balance_callback(rq, &per_cpu(dl_push_head, rq->cpu), push_dl_tasks);
657 } 561 }
658 562
659 static inline void deadline_queue_pull_task(st 563 static inline void deadline_queue_pull_task(struct rq *rq)
660 { 564 {
661 queue_balance_callback(rq, &per_cpu(dl 565 queue_balance_callback(rq, &per_cpu(dl_pull_head, rq->cpu), pull_dl_task);
662 } 566 }
663 567
664 static struct rq *find_lock_later_rq(struct ta 568 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq);
665 569
666 static struct rq *dl_task_offline_migration(st 570 static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p)
667 { 571 {
668 struct rq *later_rq = NULL; 572 struct rq *later_rq = NULL;
669 struct dl_bw *dl_b; 573 struct dl_bw *dl_b;
670 574
671 later_rq = find_lock_later_rq(p, rq); 575 later_rq = find_lock_later_rq(p, rq);
672 if (!later_rq) { 576 if (!later_rq) {
673 int cpu; 577 int cpu;
674 578
675 /* 579 /*
676 * If we cannot preempt any rq 580 * If we cannot preempt any rq, fall back to pick any
677 * online CPU: 581 * online CPU:
678 */ 582 */
679 cpu = cpumask_any_and(cpu_acti 583 cpu = cpumask_any_and(cpu_active_mask, p->cpus_ptr);
680 if (cpu >= nr_cpu_ids) { 584 if (cpu >= nr_cpu_ids) {
681 /* 585 /*
682 * Failed to find any 586 * Failed to find any suitable CPU.
683 * The task will never 587 * The task will never come back!
684 */ 588 */
685 WARN_ON_ONCE(dl_bandwi !! 589 BUG_ON(dl_bandwidth_enabled());
686 590
687 /* 591 /*
688 * If admission contro 592 * If admission control is disabled we
689 * try a little harder 593 * try a little harder to let the task
690 * run. 594 * run.
691 */ 595 */
692 cpu = cpumask_any(cpu_ 596 cpu = cpumask_any(cpu_active_mask);
693 } 597 }
694 later_rq = cpu_rq(cpu); 598 later_rq = cpu_rq(cpu);
695 double_lock_balance(rq, later_ 599 double_lock_balance(rq, later_rq);
696 } 600 }
697 601
698 if (p->dl.dl_non_contending || p->dl.d 602 if (p->dl.dl_non_contending || p->dl.dl_throttled) {
699 /* 603 /*
700 * Inactive timer is armed (or 604 * Inactive timer is armed (or callback is running, but
701 * waiting for us to release r 605 * waiting for us to release rq locks). In any case, when it
702 * will fire (or continue), it 606 * will fire (or continue), it will see running_bw of this
703 * task migrated to later_rq ( 607 * task migrated to later_rq (and correctly handle it).
704 */ 608 */
705 sub_running_bw(&p->dl, &rq->dl 609 sub_running_bw(&p->dl, &rq->dl);
706 sub_rq_bw(&p->dl, &rq->dl); 610 sub_rq_bw(&p->dl, &rq->dl);
707 611
708 add_rq_bw(&p->dl, &later_rq->d 612 add_rq_bw(&p->dl, &later_rq->dl);
709 add_running_bw(&p->dl, &later_ 613 add_running_bw(&p->dl, &later_rq->dl);
710 } else { 614 } else {
711 sub_rq_bw(&p->dl, &rq->dl); 615 sub_rq_bw(&p->dl, &rq->dl);
712 add_rq_bw(&p->dl, &later_rq->d 616 add_rq_bw(&p->dl, &later_rq->dl);
713 } 617 }
714 618
715 /* 619 /*
716 * And we finally need to fix up root_ !! 620 * And we finally need to fixup root_domain(s) bandwidth accounting,
717 * since p is still hanging out in the 621 * since p is still hanging out in the old (now moved to default) root
718 * domain. 622 * domain.
719 */ 623 */
720 dl_b = &rq->rd->dl_bw; 624 dl_b = &rq->rd->dl_bw;
721 raw_spin_lock(&dl_b->lock); 625 raw_spin_lock(&dl_b->lock);
722 __dl_sub(dl_b, p->dl.dl_bw, cpumask_we 626 __dl_sub(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
723 raw_spin_unlock(&dl_b->lock); 627 raw_spin_unlock(&dl_b->lock);
724 628
725 dl_b = &later_rq->rd->dl_bw; 629 dl_b = &later_rq->rd->dl_bw;
726 raw_spin_lock(&dl_b->lock); 630 raw_spin_lock(&dl_b->lock);
727 __dl_add(dl_b, p->dl.dl_bw, cpumask_we 631 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(later_rq->rd->span));
728 raw_spin_unlock(&dl_b->lock); 632 raw_spin_unlock(&dl_b->lock);
729 633
730 set_task_cpu(p, later_rq->cpu); 634 set_task_cpu(p, later_rq->cpu);
731 double_unlock_balance(later_rq, rq); 635 double_unlock_balance(later_rq, rq);
732 636
733 return later_rq; 637 return later_rq;
734 } 638 }
735 639
736 #else 640 #else
737 641
738 static inline 642 static inline
739 void enqueue_pushable_dl_task(struct rq *rq, s 643 void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
740 { 644 {
741 } 645 }
742 646
743 static inline 647 static inline
744 void dequeue_pushable_dl_task(struct rq *rq, s 648 void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
745 { 649 {
746 } 650 }
747 651
748 static inline 652 static inline
749 void inc_dl_migration(struct sched_dl_entity * 653 void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
750 { 654 {
751 } 655 }
752 656
753 static inline 657 static inline
754 void dec_dl_migration(struct sched_dl_entity * 658 void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
755 { 659 {
756 } 660 }
757 661
>> 662 static inline bool need_pull_dl_task(struct rq *rq, struct task_struct *prev)
>> 663 {
>> 664 return false;
>> 665 }
>> 666
>> 667 static inline void pull_dl_task(struct rq *rq)
>> 668 {
>> 669 }
>> 670
758 static inline void deadline_queue_push_tasks(s 671 static inline void deadline_queue_push_tasks(struct rq *rq)
759 { 672 {
760 } 673 }
761 674
762 static inline void deadline_queue_pull_task(st 675 static inline void deadline_queue_pull_task(struct rq *rq)
763 { 676 {
764 } 677 }
765 #endif /* CONFIG_SMP */ 678 #endif /* CONFIG_SMP */
766 679
767 static void <<
768 enqueue_dl_entity(struct sched_dl_entity *dl_s <<
769 static void enqueue_task_dl(struct rq *rq, str 680 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
770 static void dequeue_dl_entity(struct sched_dl_ !! 681 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
771 static void wakeup_preempt_dl(struct rq *rq, s !! 682 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p, int flags);
772 <<
773 static inline void replenish_dl_new_period(str <<
774 st <<
775 { <<
776 /* for non-boosted task, pi_of(dl_se) <<
777 dl_se->deadline = rq_clock(rq) + pi_of <<
778 dl_se->runtime = pi_of(dl_se)->dl_runt <<
779 <<
780 /* <<
781 * If it is a deferred reservation, an <<
782 * is not handling an starvation case, <<
783 */ <<
784 if (dl_se->dl_defer & !dl_se->dl_defer <<
785 dl_se->dl_throttled = 1; <<
786 dl_se->dl_defer_armed = 1; <<
787 } <<
788 } <<
789 683
790 /* 684 /*
791 * We are being explicitly informed that a new 685 * We are being explicitly informed that a new instance is starting,
792 * and this means that: 686 * and this means that:
793 * - the absolute deadline of the entity has 687 * - the absolute deadline of the entity has to be placed at
794 * current time + relative deadline; 688 * current time + relative deadline;
795 * - the runtime of the entity has to be set 689 * - the runtime of the entity has to be set to the maximum value.
796 * 690 *
797 * The capability of specifying such event is 691 * The capability of specifying such event is useful whenever a -deadline
798 * entity wants to (try to!) synchronize its b 692 * entity wants to (try to!) synchronize its behaviour with the scheduler's
799 * one, and to (try to!) reconcile itself with 693 * one, and to (try to!) reconcile itself with its own scheduling
800 * parameters. 694 * parameters.
801 */ 695 */
802 static inline void setup_new_dl_entity(struct 696 static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
803 { 697 {
804 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 698 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
805 struct rq *rq = rq_of_dl_rq(dl_rq); 699 struct rq *rq = rq_of_dl_rq(dl_rq);
806 700
807 WARN_ON(is_dl_boosted(dl_se)); !! 701 WARN_ON(dl_se->dl_boosted);
808 WARN_ON(dl_time_before(rq_clock(rq), d 702 WARN_ON(dl_time_before(rq_clock(rq), dl_se->deadline));
809 703
810 /* 704 /*
811 * We are racing with the deadline tim 705 * We are racing with the deadline timer. So, do nothing because
812 * the deadline timer handler will tak 706 * the deadline timer handler will take care of properly recharging
813 * the runtime and postponing the dead 707 * the runtime and postponing the deadline
814 */ 708 */
815 if (dl_se->dl_throttled) 709 if (dl_se->dl_throttled)
816 return; 710 return;
817 711
818 /* 712 /*
819 * We use the regular wall clock time 713 * We use the regular wall clock time to set deadlines in the
820 * future; in fact, we must consider e 714 * future; in fact, we must consider execution overheads (time
821 * spent on hardirq context, etc.). 715 * spent on hardirq context, etc.).
822 */ 716 */
823 replenish_dl_new_period(dl_se, rq); !! 717 dl_se->deadline = rq_clock(rq) + dl_se->dl_deadline;
>> 718 dl_se->runtime = dl_se->dl_runtime;
824 } 719 }
825 720
826 static int start_dl_timer(struct sched_dl_enti <<
827 static bool dl_entity_overflow(struct sched_dl <<
828 <<
829 /* 721 /*
830 * Pure Earliest Deadline First (EDF) scheduli 722 * Pure Earliest Deadline First (EDF) scheduling does not deal with the
831 * possibility of a entity lasting more than w 723 * possibility of a entity lasting more than what it declared, and thus
832 * exhausting its runtime. 724 * exhausting its runtime.
833 * 725 *
834 * Here we are interested in making runtime ov 726 * Here we are interested in making runtime overrun possible, but we do
835 * not want a entity which is misbehaving to a 727 * not want a entity which is misbehaving to affect the scheduling of all
836 * other entities. 728 * other entities.
837 * Therefore, a budgeting strategy called Cons 729 * Therefore, a budgeting strategy called Constant Bandwidth Server (CBS)
838 * is used, in order to confine each entity wi 730 * is used, in order to confine each entity within its own bandwidth.
839 * 731 *
840 * This function deals exactly with that, and 732 * This function deals exactly with that, and ensures that when the runtime
841 * of a entity is replenished, its deadline is 733 * of a entity is replenished, its deadline is also postponed. That ensures
842 * the overrunning entity can't interfere with 734 * the overrunning entity can't interfere with other entity in the system and
843 * can't make them miss their deadlines. Reaso 735 * can't make them miss their deadlines. Reasons why this kind of overruns
844 * could happen are, typically, a entity volun 736 * could happen are, typically, a entity voluntarily trying to overcome its
845 * runtime, or it just underestimated it durin 737 * runtime, or it just underestimated it during sched_setattr().
846 */ 738 */
847 static void replenish_dl_entity(struct sched_d !! 739 static void replenish_dl_entity(struct sched_dl_entity *dl_se,
>> 740 struct sched_dl_entity *pi_se)
848 { 741 {
849 struct dl_rq *dl_rq = dl_rq_of_se(dl_s 742 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
850 struct rq *rq = rq_of_dl_rq(dl_rq); 743 struct rq *rq = rq_of_dl_rq(dl_rq);
851 744
852 WARN_ON_ONCE(pi_of(dl_se)->dl_runtime !! 745 BUG_ON(pi_se->dl_runtime <= 0);
853 746
854 /* 747 /*
855 * This could be the case for a !-dl t 748 * This could be the case for a !-dl task that is boosted.
856 * Just go with full inherited paramet 749 * Just go with full inherited parameters.
857 * <<
858 * Or, it could be the case of a defer <<
859 * was not able to consume its runtime <<
860 * reached this point with current u > <<
861 * <<
862 * In both cases, set a new period. <<
863 */ 750 */
864 if (dl_se->dl_deadline == 0 || !! 751 if (dl_se->dl_deadline == 0) {
865 (dl_se->dl_defer_armed && dl_entit !! 752 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
866 dl_se->deadline = rq_clock(rq) !! 753 dl_se->runtime = pi_se->dl_runtime;
867 dl_se->runtime = pi_of(dl_se)- <<
868 } 754 }
869 755
870 if (dl_se->dl_yielded && dl_se->runtim 756 if (dl_se->dl_yielded && dl_se->runtime > 0)
871 dl_se->runtime = 0; 757 dl_se->runtime = 0;
872 758
873 /* 759 /*
874 * We keep moving the deadline away un 760 * We keep moving the deadline away until we get some
875 * available runtime for the entity. T 761 * available runtime for the entity. This ensures correct
876 * handling of situations where the ru 762 * handling of situations where the runtime overrun is
877 * arbitrary large. 763 * arbitrary large.
878 */ 764 */
879 while (dl_se->runtime <= 0) { 765 while (dl_se->runtime <= 0) {
880 dl_se->deadline += pi_of(dl_se !! 766 dl_se->deadline += pi_se->dl_period;
881 dl_se->runtime += pi_of(dl_se) !! 767 dl_se->runtime += pi_se->dl_runtime;
882 } 768 }
883 769
884 /* 770 /*
885 * At this point, the deadline really 771 * At this point, the deadline really should be "in
886 * the future" with respect to rq->clo 772 * the future" with respect to rq->clock. If it's
887 * not, we are, for some reason, laggi 773 * not, we are, for some reason, lagging too much!
888 * Anyway, after having warn userspace 774 * Anyway, after having warn userspace abut that,
889 * we still try to keep the things run 775 * we still try to keep the things running by
890 * resetting the deadline and the budg 776 * resetting the deadline and the budget of the
891 * entity. 777 * entity.
892 */ 778 */
893 if (dl_time_before(dl_se->deadline, rq 779 if (dl_time_before(dl_se->deadline, rq_clock(rq))) {
894 printk_deferred_once("sched: D 780 printk_deferred_once("sched: DL replenish lagged too much\n");
895 replenish_dl_new_period(dl_se, !! 781 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
>> 782 dl_se->runtime = pi_se->dl_runtime;
896 } 783 }
897 784
898 if (dl_se->dl_yielded) 785 if (dl_se->dl_yielded)
899 dl_se->dl_yielded = 0; 786 dl_se->dl_yielded = 0;
900 if (dl_se->dl_throttled) 787 if (dl_se->dl_throttled)
901 dl_se->dl_throttled = 0; 788 dl_se->dl_throttled = 0;
902 <<
903 /* <<
904 * If this is the replenishment of a d <<
905 * clear the flag and return. <<
906 */ <<
907 if (dl_se->dl_defer_armed) { <<
908 dl_se->dl_defer_armed = 0; <<
909 return; <<
910 } <<
911 <<
912 /* <<
913 * A this point, if the deferred serve <<
914 * is in the future, if it is not runn <<
915 * and arm the defer timer. <<
916 */ <<
917 if (dl_se->dl_defer && !dl_se->dl_defe <<
918 dl_time_before(rq_clock(dl_se->rq) <<
919 if (!is_dl_boosted(dl_se) && d <<
920 <<
921 /* <<
922 * Set dl_se->dl_defer <<
923 * inform the start_dl <<
924 * activation. <<
925 */ <<
926 dl_se->dl_defer_armed <<
927 dl_se->dl_throttled = <<
928 if (!start_dl_timer(dl <<
929 /* <<
930 * If for what <<
931 * queued but <<
932 * deferrable <<
933 */ <<
934 hrtimer_try_to <<
935 dl_se->dl_defe <<
936 dl_se->dl_thro <<
937 } <<
938 } <<
939 } <<
940 } 789 }
941 790
942 /* 791 /*
943 * Here we check if --at time t-- an entity (w 792 * Here we check if --at time t-- an entity (which is probably being
944 * [re]activated or, in general, enqueued) can 793 * [re]activated or, in general, enqueued) can use its remaining runtime
945 * and its current deadline _without_ exceedin 794 * and its current deadline _without_ exceeding the bandwidth it is
946 * assigned (function returns true if it can't 795 * assigned (function returns true if it can't). We are in fact applying
947 * one of the CBS rules: when a task wakes up, 796 * one of the CBS rules: when a task wakes up, if the residual runtime
948 * over residual deadline fits within the allo 797 * over residual deadline fits within the allocated bandwidth, then we
949 * can keep the current (absolute) deadline an 798 * can keep the current (absolute) deadline and residual budget without
950 * disrupting the schedulability of the system 799 * disrupting the schedulability of the system. Otherwise, we should
951 * refill the runtime and set the deadline a p 800 * refill the runtime and set the deadline a period in the future,
952 * because keeping the current (absolute) dead 801 * because keeping the current (absolute) deadline of the task would
953 * result in breaking guarantees promised to o 802 * result in breaking guarantees promised to other tasks (refer to
954 * Documentation/scheduler/sched-deadline.rst 803 * Documentation/scheduler/sched-deadline.rst for more information).
955 * 804 *
956 * This function returns true if: 805 * This function returns true if:
957 * 806 *
958 * runtime / (deadline - t) > dl_runtime / d 807 * runtime / (deadline - t) > dl_runtime / dl_deadline ,
959 * 808 *
960 * IOW we can't recycle current parameters. 809 * IOW we can't recycle current parameters.
961 * 810 *
962 * Notice that the bandwidth check is done aga 811 * Notice that the bandwidth check is done against the deadline. For
963 * task with deadline equal to period this is 812 * task with deadline equal to period this is the same of using
964 * dl_period instead of dl_deadline in the equ 813 * dl_period instead of dl_deadline in the equation above.
965 */ 814 */
966 static bool dl_entity_overflow(struct sched_dl !! 815 static bool dl_entity_overflow(struct sched_dl_entity *dl_se,
>> 816 struct sched_dl_entity *pi_se, u64 t)
967 { 817 {
968 u64 left, right; 818 u64 left, right;
969 819
970 /* 820 /*
971 * left and right are the two sides of 821 * left and right are the two sides of the equation above,
972 * after a bit of shuffling to use mul 822 * after a bit of shuffling to use multiplications instead
973 * of divisions. 823 * of divisions.
974 * 824 *
975 * Note that none of the time values i 825 * Note that none of the time values involved in the two
976 * multiplications are absolute: dl_de 826 * multiplications are absolute: dl_deadline and dl_runtime
977 * are the relative deadline and the m 827 * are the relative deadline and the maximum runtime of each
978 * instance, runtime is the runtime le 828 * instance, runtime is the runtime left for the last instance
979 * and (deadline - t), since t is rq-> 829 * and (deadline - t), since t is rq->clock, is the time left
980 * to the (absolute) deadline. Even if 830 * to the (absolute) deadline. Even if overflowing the u64 type
981 * is very unlikely to occur in both c 831 * is very unlikely to occur in both cases, here we scale down
982 * as we want to avoid that risk at al 832 * as we want to avoid that risk at all. Scaling down by 10
983 * means that we reduce granularity to 833 * means that we reduce granularity to 1us. We are fine with it,
984 * since this is only a true/false che 834 * since this is only a true/false check and, anyway, thinking
985 * of anything below microseconds reso 835 * of anything below microseconds resolution is actually fiction
986 * (but still we want to give the user 836 * (but still we want to give the user that illusion >;).
987 */ 837 */
988 left = (pi_of(dl_se)->dl_deadline >> D !! 838 left = (pi_se->dl_deadline >> DL_SCALE) * (dl_se->runtime >> DL_SCALE);
989 right = ((dl_se->deadline - t) >> DL_S 839 right = ((dl_se->deadline - t) >> DL_SCALE) *
990 (pi_of(dl_se)->dl_runtime >> D !! 840 (pi_se->dl_runtime >> DL_SCALE);
991 841
992 return dl_time_before(right, left); 842 return dl_time_before(right, left);
993 } 843 }
994 844
995 /* 845 /*
996 * Revised wakeup rule [1]: For self-suspendin 846 * Revised wakeup rule [1]: For self-suspending tasks, rather then
997 * re-initializing task's runtime and deadline 847 * re-initializing task's runtime and deadline, the revised wakeup
998 * rule adjusts the task's runtime to avoid th 848 * rule adjusts the task's runtime to avoid the task to overrun its
999 * density. 849 * density.
1000 * 850 *
1001 * Reasoning: a task may overrun the density 851 * Reasoning: a task may overrun the density if:
1002 * runtime / (deadline - t) > dl_runtime / 852 * runtime / (deadline - t) > dl_runtime / dl_deadline
1003 * 853 *
1004 * Therefore, runtime can be adjusted to: 854 * Therefore, runtime can be adjusted to:
1005 * runtime = (dl_runtime / dl_deadline) * 855 * runtime = (dl_runtime / dl_deadline) * (deadline - t)
1006 * 856 *
1007 * In such way that runtime will be equal to 857 * In such way that runtime will be equal to the maximum density
1008 * the task can use without breaking any rule 858 * the task can use without breaking any rule.
1009 * 859 *
1010 * [1] Luca Abeni, Giuseppe Lipari, and Juri 860 * [1] Luca Abeni, Giuseppe Lipari, and Juri Lelli. 2015. Constant
1011 * bandwidth server revisited. SIGBED Rev. 11 861 * bandwidth server revisited. SIGBED Rev. 11, 4 (January 2015), 19-24.
1012 */ 862 */
1013 static void 863 static void
1014 update_dl_revised_wakeup(struct sched_dl_enti 864 update_dl_revised_wakeup(struct sched_dl_entity *dl_se, struct rq *rq)
1015 { 865 {
1016 u64 laxity = dl_se->deadline - rq_clo 866 u64 laxity = dl_se->deadline - rq_clock(rq);
1017 867
1018 /* 868 /*
1019 * If the task has deadline < period, 869 * If the task has deadline < period, and the deadline is in the past,
1020 * it should already be throttled bef 870 * it should already be throttled before this check.
1021 * 871 *
1022 * See update_dl_entity() comments fo 872 * See update_dl_entity() comments for further details.
1023 */ 873 */
1024 WARN_ON(dl_time_before(dl_se->deadlin 874 WARN_ON(dl_time_before(dl_se->deadline, rq_clock(rq)));
1025 875
1026 dl_se->runtime = (dl_se->dl_density * 876 dl_se->runtime = (dl_se->dl_density * laxity) >> BW_SHIFT;
1027 } 877 }
1028 878
1029 /* 879 /*
1030 * Regarding the deadline, a task with implic 880 * Regarding the deadline, a task with implicit deadline has a relative
1031 * deadline == relative period. A task with c 881 * deadline == relative period. A task with constrained deadline has a
1032 * relative deadline <= relative period. 882 * relative deadline <= relative period.
1033 * 883 *
1034 * We support constrained deadline tasks. How 884 * We support constrained deadline tasks. However, there are some restrictions
1035 * applied only for tasks which do not have a 885 * applied only for tasks which do not have an implicit deadline. See
1036 * update_dl_entity() to know more about such 886 * update_dl_entity() to know more about such restrictions.
1037 * 887 *
1038 * The dl_is_implicit() returns true if the t 888 * The dl_is_implicit() returns true if the task has an implicit deadline.
1039 */ 889 */
1040 static inline bool dl_is_implicit(struct sche 890 static inline bool dl_is_implicit(struct sched_dl_entity *dl_se)
1041 { 891 {
1042 return dl_se->dl_deadline == dl_se->d 892 return dl_se->dl_deadline == dl_se->dl_period;
1043 } 893 }
1044 894
1045 /* 895 /*
1046 * When a deadline entity is placed in the ru 896 * When a deadline entity is placed in the runqueue, its runtime and deadline
1047 * might need to be updated. This is done by 897 * might need to be updated. This is done by a CBS wake up rule. There are two
1048 * different rules: 1) the original CBS; and 898 * different rules: 1) the original CBS; and 2) the Revisited CBS.
1049 * 899 *
1050 * When the task is starting a new period, th 900 * When the task is starting a new period, the Original CBS is used. In this
1051 * case, the runtime is replenished and a new 901 * case, the runtime is replenished and a new absolute deadline is set.
1052 * 902 *
1053 * When a task is queued before the begin of 903 * When a task is queued before the begin of the next period, using the
1054 * remaining runtime and deadline could make 904 * remaining runtime and deadline could make the entity to overflow, see
1055 * dl_entity_overflow() to find more about ru 905 * dl_entity_overflow() to find more about runtime overflow. When such case
1056 * is detected, the runtime and deadline need 906 * is detected, the runtime and deadline need to be updated.
1057 * 907 *
1058 * If the task has an implicit deadline, i.e. 908 * If the task has an implicit deadline, i.e., deadline == period, the Original
1059 * CBS is applied. The runtime is replenished !! 909 * CBS is applied. the runtime is replenished and a new absolute deadline is
1060 * set, as in the previous cases. 910 * set, as in the previous cases.
1061 * 911 *
1062 * However, the Original CBS does not work pr 912 * However, the Original CBS does not work properly for tasks with
1063 * deadline < period, which are said to have 913 * deadline < period, which are said to have a constrained deadline. By
1064 * applying the Original CBS, a constrained d 914 * applying the Original CBS, a constrained deadline task would be able to run
1065 * runtime/deadline in a period. With deadlin 915 * runtime/deadline in a period. With deadline < period, the task would
1066 * overrun the runtime/period allowed bandwid 916 * overrun the runtime/period allowed bandwidth, breaking the admission test.
1067 * 917 *
1068 * In order to prevent this misbehave, the Re 918 * In order to prevent this misbehave, the Revisited CBS is used for
1069 * constrained deadline tasks when a runtime 919 * constrained deadline tasks when a runtime overflow is detected. In the
1070 * Revisited CBS, rather than replenishing & 920 * Revisited CBS, rather than replenishing & setting a new absolute deadline,
1071 * the remaining runtime of the task is reduc 921 * the remaining runtime of the task is reduced to avoid runtime overflow.
1072 * Please refer to the comments update_dl_rev 922 * Please refer to the comments update_dl_revised_wakeup() function to find
1073 * more about the Revised CBS rule. 923 * more about the Revised CBS rule.
1074 */ 924 */
1075 static void update_dl_entity(struct sched_dl_ !! 925 static void update_dl_entity(struct sched_dl_entity *dl_se,
>> 926 struct sched_dl_entity *pi_se)
1076 { 927 {
1077 struct rq *rq = rq_of_dl_se(dl_se); !! 928 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
>> 929 struct rq *rq = rq_of_dl_rq(dl_rq);
1078 930
1079 if (dl_time_before(dl_se->deadline, r 931 if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
1080 dl_entity_overflow(dl_se, rq_cloc !! 932 dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
1081 933
1082 if (unlikely(!dl_is_implicit( 934 if (unlikely(!dl_is_implicit(dl_se) &&
1083 !dl_time_before( 935 !dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1084 !is_dl_boosted(d !! 936 !dl_se->dl_boosted)){
1085 update_dl_revised_wak 937 update_dl_revised_wakeup(dl_se, rq);
1086 return; 938 return;
1087 } 939 }
1088 940
1089 replenish_dl_new_period(dl_se !! 941 dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
1090 } else if (dl_server(dl_se) && dl_se- !! 942 dl_se->runtime = pi_se->dl_runtime;
1091 /* <<
1092 * The server can still use i <<
1093 * it left the dl_defer_runni <<
1094 */ <<
1095 if (!dl_se->dl_defer_running) <<
1096 dl_se->dl_defer_armed <<
1097 dl_se->dl_throttled = <<
1098 } <<
1099 } 943 }
1100 } 944 }
1101 945
1102 static inline u64 dl_next_period(struct sched 946 static inline u64 dl_next_period(struct sched_dl_entity *dl_se)
1103 { 947 {
1104 return dl_se->deadline - dl_se->dl_de 948 return dl_se->deadline - dl_se->dl_deadline + dl_se->dl_period;
1105 } 949 }
1106 950
1107 /* 951 /*
1108 * If the entity depleted all its runtime, an 952 * If the entity depleted all its runtime, and if we want it to sleep
1109 * while waiting for some new execution time 953 * while waiting for some new execution time to become available, we
1110 * set the bandwidth replenishment timer to t 954 * set the bandwidth replenishment timer to the replenishment instant
1111 * and try to activate it. 955 * and try to activate it.
1112 * 956 *
1113 * Notice that it is important for the caller 957 * Notice that it is important for the caller to know if the timer
1114 * actually started or not (i.e., the repleni 958 * actually started or not (i.e., the replenishment instant is in
1115 * the future or in the past). 959 * the future or in the past).
1116 */ 960 */
1117 static int start_dl_timer(struct sched_dl_ent !! 961 static int start_dl_timer(struct task_struct *p)
1118 { 962 {
>> 963 struct sched_dl_entity *dl_se = &p->dl;
1119 struct hrtimer *timer = &dl_se->dl_ti 964 struct hrtimer *timer = &dl_se->dl_timer;
1120 struct dl_rq *dl_rq = dl_rq_of_se(dl_ !! 965 struct rq *rq = task_rq(p);
1121 struct rq *rq = rq_of_dl_rq(dl_rq); <<
1122 ktime_t now, act; 966 ktime_t now, act;
1123 s64 delta; 967 s64 delta;
1124 968
1125 lockdep_assert_rq_held(rq); !! 969 lockdep_assert_held(&rq->lock);
1126 970
1127 /* 971 /*
1128 * We want the timer to fire at the d 972 * We want the timer to fire at the deadline, but considering
1129 * that it is actually coming from rq 973 * that it is actually coming from rq->clock and not from
1130 * hrtimer's time base reading. 974 * hrtimer's time base reading.
1131 * !! 975 */
1132 * The deferred reservation will have !! 976 act = ns_to_ktime(dl_next_period(dl_se));
1133 * (deadline - runtime). At that poin <<
1134 * if the current deadline can be use <<
1135 * required to avoid add too much pre <<
1136 * (current u > U). <<
1137 */ <<
1138 if (dl_se->dl_defer_armed) { <<
1139 WARN_ON_ONCE(!dl_se->dl_throt <<
1140 act = ns_to_ktime(dl_se->dead <<
1141 } else { <<
1142 /* act = deadline - rel-deadl <<
1143 act = ns_to_ktime(dl_next_per <<
1144 } <<
1145 <<
1146 now = hrtimer_cb_get_time(timer); 977 now = hrtimer_cb_get_time(timer);
1147 delta = ktime_to_ns(now) - rq_clock(r 978 delta = ktime_to_ns(now) - rq_clock(rq);
1148 act = ktime_add_ns(act, delta); 979 act = ktime_add_ns(act, delta);
1149 980
1150 /* 981 /*
1151 * If the expiry time already passed, 982 * If the expiry time already passed, e.g., because the value
1152 * chosen as the deadline is too smal 983 * chosen as the deadline is too small, don't even try to
1153 * start the timer in the past! 984 * start the timer in the past!
1154 */ 985 */
1155 if (ktime_us_delta(act, now) < 0) 986 if (ktime_us_delta(act, now) < 0)
1156 return 0; 987 return 0;
1157 988
1158 /* 989 /*
1159 * !enqueued will guarantee another c 990 * !enqueued will guarantee another callback; even if one is already in
1160 * progress. This ensures a balanced 991 * progress. This ensures a balanced {get,put}_task_struct().
1161 * 992 *
1162 * The race against __run_timer() cle 993 * The race against __run_timer() clearing the enqueued state is
1163 * harmless because we're holding tas 994 * harmless because we're holding task_rq()->lock, therefore the timer
1164 * expiring after we've done the chec 995 * expiring after we've done the check will wait on its task_rq_lock()
1165 * and observe our state. 996 * and observe our state.
1166 */ 997 */
1167 if (!hrtimer_is_queued(timer)) { 998 if (!hrtimer_is_queued(timer)) {
1168 if (!dl_server(dl_se)) !! 999 get_task_struct(p);
1169 get_task_struct(dl_ta <<
1170 hrtimer_start(timer, act, HRT 1000 hrtimer_start(timer, act, HRTIMER_MODE_ABS_HARD);
1171 } 1001 }
1172 1002
1173 return 1; 1003 return 1;
1174 } 1004 }
1175 1005
1176 static void __push_dl_task(struct rq *rq, str <<
1177 { <<
1178 #ifdef CONFIG_SMP <<
1179 /* <<
1180 * Queueing this task back might have <<
1181 * to kick someone away. <<
1182 */ <<
1183 if (has_pushable_dl_tasks(rq)) { <<
1184 /* <<
1185 * Nothing relies on rq->lock <<
1186 * rq->lock. <<
1187 */ <<
1188 rq_unpin_lock(rq, rf); <<
1189 push_dl_task(rq); <<
1190 rq_repin_lock(rq, rf); <<
1191 } <<
1192 #endif <<
1193 } <<
1194 <<
1195 /* a defer timer will not be reset if the run <<
1196 static const u64 dl_server_min_res = 1 * NSEC <<
1197 <<
1198 static enum hrtimer_restart dl_server_timer(s <<
1199 { <<
1200 struct rq *rq = rq_of_dl_se(dl_se); <<
1201 u64 fw; <<
1202 <<
1203 scoped_guard (rq_lock, rq) { <<
1204 struct rq_flags *rf = &scope. <<
1205 <<
1206 if (!dl_se->dl_throttled || ! <<
1207 return HRTIMER_NOREST <<
1208 <<
1209 sched_clock_tick(); <<
1210 update_rq_clock(rq); <<
1211 <<
1212 if (!dl_se->dl_runtime) <<
1213 return HRTIMER_NOREST <<
1214 <<
1215 if (!dl_se->server_has_tasks( <<
1216 replenish_dl_entity(d <<
1217 return HRTIMER_NOREST <<
1218 } <<
1219 <<
1220 if (dl_se->dl_defer_armed) { <<
1221 /* <<
1222 * First check if the <<
1223 * If so, it is possi <<
1224 * of time. The dl_se <<
1225 * forwarding the tim <<
1226 */ <<
1227 if (dl_time_before(rq <<
1228 (d <<
1229 <<
1230 /* reset the <<
1231 fw = dl_se->d <<
1232 <<
1233 hrtimer_forwa <<
1234 return HRTIME <<
1235 } <<
1236 <<
1237 dl_se->dl_defer_runni <<
1238 } <<
1239 <<
1240 enqueue_dl_entity(dl_se, ENQU <<
1241 <<
1242 if (!dl_task(dl_se->rq->curr) <<
1243 resched_curr(rq); <<
1244 <<
1245 __push_dl_task(rq, rf); <<
1246 } <<
1247 <<
1248 return HRTIMER_NORESTART; <<
1249 } <<
1250 <<
1251 /* 1006 /*
1252 * This is the bandwidth enforcement timer ca 1007 * This is the bandwidth enforcement timer callback. If here, we know
1253 * a task is not on its dl_rq, since the fact 1008 * a task is not on its dl_rq, since the fact that the timer was running
1254 * means the task is throttled and needs a ru 1009 * means the task is throttled and needs a runtime replenishment.
1255 * 1010 *
1256 * However, what we actually do depends on th 1011 * However, what we actually do depends on the fact the task is active,
1257 * (it is on its rq) or has been removed from 1012 * (it is on its rq) or has been removed from there by a call to
1258 * dequeue_task_dl(). In the former case we m 1013 * dequeue_task_dl(). In the former case we must issue the runtime
1259 * replenishment and add the task back to the 1014 * replenishment and add the task back to the dl_rq; in the latter, we just
1260 * do nothing but clearing dl_throttled, so t 1015 * do nothing but clearing dl_throttled, so that runtime and deadline
1261 * updating (and the queueing back to dl_rq) 1016 * updating (and the queueing back to dl_rq) will be done by the
1262 * next call to enqueue_task_dl(). 1017 * next call to enqueue_task_dl().
1263 */ 1018 */
1264 static enum hrtimer_restart dl_task_timer(str 1019 static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
1265 { 1020 {
1266 struct sched_dl_entity *dl_se = conta 1021 struct sched_dl_entity *dl_se = container_of(timer,
1267 1022 struct sched_dl_entity,
1268 1023 dl_timer);
1269 struct task_struct *p; !! 1024 struct task_struct *p = dl_task_of(dl_se);
1270 struct rq_flags rf; 1025 struct rq_flags rf;
1271 struct rq *rq; 1026 struct rq *rq;
1272 1027
1273 if (dl_server(dl_se)) <<
1274 return dl_server_timer(timer, <<
1275 <<
1276 p = dl_task_of(dl_se); <<
1277 rq = task_rq_lock(p, &rf); 1028 rq = task_rq_lock(p, &rf);
1278 1029
1279 /* 1030 /*
1280 * The task might have changed its sc 1031 * The task might have changed its scheduling policy to something
1281 * different than SCHED_DEADLINE (thr 1032 * different than SCHED_DEADLINE (through switched_from_dl()).
1282 */ 1033 */
1283 if (!dl_task(p)) 1034 if (!dl_task(p))
1284 goto unlock; 1035 goto unlock;
1285 1036
1286 /* 1037 /*
1287 * The task might have been boosted b 1038 * The task might have been boosted by someone else and might be in the
1288 * boosting/deboosting path, its not 1039 * boosting/deboosting path, its not throttled.
1289 */ 1040 */
1290 if (is_dl_boosted(dl_se)) !! 1041 if (dl_se->dl_boosted)
1291 goto unlock; 1042 goto unlock;
1292 1043
1293 /* 1044 /*
1294 * Spurious timer due to start_dl_tim 1045 * Spurious timer due to start_dl_timer() race; or we already received
1295 * a replenishment from rt_mutex_setp 1046 * a replenishment from rt_mutex_setprio().
1296 */ 1047 */
1297 if (!dl_se->dl_throttled) 1048 if (!dl_se->dl_throttled)
1298 goto unlock; 1049 goto unlock;
1299 1050
1300 sched_clock_tick(); 1051 sched_clock_tick();
1301 update_rq_clock(rq); 1052 update_rq_clock(rq);
1302 1053
1303 /* 1054 /*
1304 * If the throttle happened during sc 1055 * If the throttle happened during sched-out; like:
1305 * 1056 *
1306 * schedule() 1057 * schedule()
1307 * deactivate_task() 1058 * deactivate_task()
1308 * dequeue_task_dl() 1059 * dequeue_task_dl()
1309 * update_curr_dl() 1060 * update_curr_dl()
1310 * start_dl_timer() 1061 * start_dl_timer()
1311 * __dequeue_task_dl() 1062 * __dequeue_task_dl()
1312 * prev->on_rq = 0; 1063 * prev->on_rq = 0;
1313 * 1064 *
1314 * We can be both throttled and !queu 1065 * We can be both throttled and !queued. Replenish the counter
1315 * but do not enqueue -- wait for our 1066 * but do not enqueue -- wait for our wakeup to do that.
1316 */ 1067 */
1317 if (!task_on_rq_queued(p)) { 1068 if (!task_on_rq_queued(p)) {
1318 replenish_dl_entity(dl_se); !! 1069 replenish_dl_entity(dl_se, dl_se);
1319 goto unlock; 1070 goto unlock;
1320 } 1071 }
1321 1072
1322 #ifdef CONFIG_SMP 1073 #ifdef CONFIG_SMP
1323 if (unlikely(!rq->online)) { 1074 if (unlikely(!rq->online)) {
1324 /* 1075 /*
1325 * If the runqueue is no long 1076 * If the runqueue is no longer available, migrate the
1326 * task elsewhere. This neces 1077 * task elsewhere. This necessarily changes rq.
1327 */ 1078 */
1328 lockdep_unpin_lock(__rq_lockp !! 1079 lockdep_unpin_lock(&rq->lock, rf.cookie);
1329 rq = dl_task_offline_migratio 1080 rq = dl_task_offline_migration(rq, p);
1330 rf.cookie = lockdep_pin_lock( !! 1081 rf.cookie = lockdep_pin_lock(&rq->lock);
1331 update_rq_clock(rq); 1082 update_rq_clock(rq);
1332 1083
1333 /* 1084 /*
1334 * Now that the task has been 1085 * Now that the task has been migrated to the new RQ and we
1335 * have that locked, proceed 1086 * have that locked, proceed as normal and enqueue the task
1336 * there. 1087 * there.
1337 */ 1088 */
1338 } 1089 }
1339 #endif 1090 #endif
1340 1091
1341 enqueue_task_dl(rq, p, ENQUEUE_REPLEN 1092 enqueue_task_dl(rq, p, ENQUEUE_REPLENISH);
1342 if (dl_task(rq->curr)) 1093 if (dl_task(rq->curr))
1343 wakeup_preempt_dl(rq, p, 0); !! 1094 check_preempt_curr_dl(rq, p, 0);
1344 else 1095 else
1345 resched_curr(rq); 1096 resched_curr(rq);
1346 1097
1347 __push_dl_task(rq, &rf); !! 1098 #ifdef CONFIG_SMP
>> 1099 /*
>> 1100 * Queueing this task back might have overloaded rq, check if we need
>> 1101 * to kick someone away.
>> 1102 */
>> 1103 if (has_pushable_dl_tasks(rq)) {
>> 1104 /*
>> 1105 * Nothing relies on rq->lock after this, so its safe to drop
>> 1106 * rq->lock.
>> 1107 */
>> 1108 rq_unpin_lock(rq, &rf);
>> 1109 push_dl_task(rq);
>> 1110 rq_repin_lock(rq, &rf);
>> 1111 }
>> 1112 #endif
1348 1113
1349 unlock: 1114 unlock:
1350 task_rq_unlock(rq, p, &rf); 1115 task_rq_unlock(rq, p, &rf);
1351 1116
1352 /* 1117 /*
1353 * This can free the task_struct, inc 1118 * This can free the task_struct, including this hrtimer, do not touch
1354 * anything related to that after thi 1119 * anything related to that after this.
1355 */ 1120 */
1356 put_task_struct(p); 1121 put_task_struct(p);
1357 1122
1358 return HRTIMER_NORESTART; 1123 return HRTIMER_NORESTART;
1359 } 1124 }
1360 1125
1361 static void init_dl_task_timer(struct sched_d !! 1126 void init_dl_task_timer(struct sched_dl_entity *dl_se)
1362 { 1127 {
1363 struct hrtimer *timer = &dl_se->dl_ti 1128 struct hrtimer *timer = &dl_se->dl_timer;
1364 1129
1365 hrtimer_init(timer, CLOCK_MONOTONIC, 1130 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1366 timer->function = dl_task_timer; 1131 timer->function = dl_task_timer;
1367 } 1132 }
1368 1133
1369 /* 1134 /*
1370 * During the activation, CBS checks if it ca 1135 * During the activation, CBS checks if it can reuse the current task's
1371 * runtime and period. If the deadline of the 1136 * runtime and period. If the deadline of the task is in the past, CBS
1372 * cannot use the runtime, and so it replenis 1137 * cannot use the runtime, and so it replenishes the task. This rule
1373 * works fine for implicit deadline tasks (de 1138 * works fine for implicit deadline tasks (deadline == period), and the
1374 * CBS was designed for implicit deadline tas 1139 * CBS was designed for implicit deadline tasks. However, a task with
1375 * constrained deadline (deadline < period) m 1140 * constrained deadline (deadline < period) might be awakened after the
1376 * deadline, but before the next period. In t 1141 * deadline, but before the next period. In this case, replenishing the
1377 * task would allow it to run for runtime / d 1142 * task would allow it to run for runtime / deadline. As in this case
1378 * deadline < period, CBS enables a task to r 1143 * deadline < period, CBS enables a task to run for more than the
1379 * runtime / period. In a very loaded system, 1144 * runtime / period. In a very loaded system, this can cause a domino
1380 * effect, making other tasks miss their dead 1145 * effect, making other tasks miss their deadlines.
1381 * 1146 *
1382 * To avoid this problem, in the activation o 1147 * To avoid this problem, in the activation of a constrained deadline
1383 * task after the deadline but before the nex 1148 * task after the deadline but before the next period, throttle the
1384 * task and set the replenishing timer to the 1149 * task and set the replenishing timer to the begin of the next period,
1385 * unless it is boosted. 1150 * unless it is boosted.
1386 */ 1151 */
1387 static inline void dl_check_constrained_dl(st 1152 static inline void dl_check_constrained_dl(struct sched_dl_entity *dl_se)
1388 { 1153 {
1389 struct rq *rq = rq_of_dl_se(dl_se); !! 1154 struct task_struct *p = dl_task_of(dl_se);
>> 1155 struct rq *rq = rq_of_dl_rq(dl_rq_of_se(dl_se));
1390 1156
1391 if (dl_time_before(dl_se->deadline, r 1157 if (dl_time_before(dl_se->deadline, rq_clock(rq)) &&
1392 dl_time_before(rq_clock(rq), dl_n 1158 dl_time_before(rq_clock(rq), dl_next_period(dl_se))) {
1393 if (unlikely(is_dl_boosted(dl !! 1159 if (unlikely(dl_se->dl_boosted || !start_dl_timer(p)))
1394 return; 1160 return;
1395 dl_se->dl_throttled = 1; 1161 dl_se->dl_throttled = 1;
1396 if (dl_se->runtime > 0) 1162 if (dl_se->runtime > 0)
1397 dl_se->runtime = 0; 1163 dl_se->runtime = 0;
1398 } 1164 }
1399 } 1165 }
1400 1166
1401 static 1167 static
1402 int dl_runtime_exceeded(struct sched_dl_entit 1168 int dl_runtime_exceeded(struct sched_dl_entity *dl_se)
1403 { 1169 {
1404 return (dl_se->runtime <= 0); 1170 return (dl_se->runtime <= 0);
1405 } 1171 }
1406 1172
>> 1173 extern bool sched_rt_bandwidth_account(struct rt_rq *rt_rq);
>> 1174
1407 /* 1175 /*
1408 * This function implements the GRUB accounti !! 1176 * This function implements the GRUB accounting rule:
1409 * GRUB reclaiming algorithm, the runtime is !! 1177 * according to the GRUB reclaiming algorithm, the runtime is
1410 * but as "dq = -(max{u, (Umax - Uinact - Uex !! 1178 * not decreased as "dq = -dt", but as
>> 1179 * "dq = -max{u / Umax, (1 - Uinact - Uextra)} dt",
1411 * where u is the utilization of the task, Um 1180 * where u is the utilization of the task, Umax is the maximum reclaimable
1412 * utilization, Uinact is the (per-runqueue) 1181 * utilization, Uinact is the (per-runqueue) inactive utilization, computed
1413 * as the difference between the "total runqu 1182 * as the difference between the "total runqueue utilization" and the
1414 * "runqueue active utilization", and Uextra !! 1183 * runqueue active utilization, and Uextra is the (per runqueue) extra
1415 * reclaimable utilization. 1184 * reclaimable utilization.
1416 * Since rq->dl.running_bw and rq->dl.this_bw !! 1185 * Since rq->dl.running_bw and rq->dl.this_bw contain utilizations
1417 * by 2^BW_SHIFT, the result has to be shifte !! 1186 * multiplied by 2^BW_SHIFT, the result has to be shifted right by
1418 * Since rq->dl.bw_ratio contains 1 / Umax mu !! 1187 * BW_SHIFT.
1419 * is multiplied by rq->dl.bw_ratio and shift !! 1188 * Since rq->dl.bw_ratio contains 1 / Umax multipled by 2^RATIO_SHIFT,
1420 * Since delta is a 64 bit variable, to have !! 1189 * dl_bw is multiped by rq->dl.bw_ratio and shifted right by RATIO_SHIFT.
1421 * larger than 2^(64 - 20 - 8), which is more !! 1190 * Since delta is a 64 bit variable, to have an overflow its value
1422 * not an issue here. !! 1191 * should be larger than 2^(64 - 20 - 8), which is more than 64 seconds.
>> 1192 * So, overflow is not an issue here.
1423 */ 1193 */
1424 static u64 grub_reclaim(u64 delta, struct rq 1194 static u64 grub_reclaim(u64 delta, struct rq *rq, struct sched_dl_entity *dl_se)
1425 { 1195 {
1426 u64 u_act; <<
1427 u64 u_inact = rq->dl.this_bw - rq->dl 1196 u64 u_inact = rq->dl.this_bw - rq->dl.running_bw; /* Utot - Uact */
>> 1197 u64 u_act;
>> 1198 u64 u_act_min = (dl_se->dl_bw * rq->dl.bw_ratio) >> RATIO_SHIFT;
1428 1199
1429 /* 1200 /*
1430 * Instead of computing max{u, (u_max !! 1201 * Instead of computing max{u * bw_ratio, (1 - u_inact - u_extra)},
1431 * compare u_inact + u_extra with u_m !! 1202 * we compare u_inact + rq->dl.extra_bw with
1432 * can be larger than u_max. So, u_ma !! 1203 * 1 - (u * rq->dl.bw_ratio >> RATIO_SHIFT), because
1433 * negative leading to wrong results. !! 1204 * u_inact + rq->dl.extra_bw can be larger than
>> 1205 * 1 * (so, 1 - u_inact - rq->dl.extra_bw would be negative
>> 1206 * leading to wrong results)
1434 */ 1207 */
1435 if (u_inact + rq->dl.extra_bw > rq->d !! 1208 if (u_inact + rq->dl.extra_bw > BW_UNIT - u_act_min)
1436 u_act = dl_se->dl_bw; !! 1209 u_act = u_act_min;
1437 else 1210 else
1438 u_act = rq->dl.max_bw - u_ina !! 1211 u_act = BW_UNIT - u_inact - rq->dl.extra_bw;
1439 1212
1440 u_act = (u_act * rq->dl.bw_ratio) >> <<
1441 return (delta * u_act) >> BW_SHIFT; 1213 return (delta * u_act) >> BW_SHIFT;
1442 } 1214 }
1443 1215
1444 s64 dl_scaled_delta_exec(struct rq *rq, struc !! 1216 /*
>> 1217 * Update the current task's runtime statistics (provided it is still
>> 1218 * a -deadline task and has not been removed from the dl_rq).
>> 1219 */
>> 1220 static void update_curr_dl(struct rq *rq)
1445 { 1221 {
1446 s64 scaled_delta_exec; !! 1222 struct task_struct *curr = rq->curr;
>> 1223 struct sched_dl_entity *dl_se = &curr->dl;
>> 1224 u64 delta_exec, scaled_delta_exec;
>> 1225 int cpu = cpu_of(rq);
>> 1226 u64 now;
>> 1227
>> 1228 if (!dl_task(curr) || !on_dl_rq(dl_se))
>> 1229 return;
>> 1230
>> 1231 /*
>> 1232 * Consumed budget is computed considering the time as
>> 1233 * observed by schedulable tasks (excluding time spent
>> 1234 * in hardirq context, etc.). Deadlines are instead
>> 1235 * computed using hard walltime. This seems to be the more
>> 1236 * natural solution, but the full ramifications of this
>> 1237 * approach need further study.
>> 1238 */
>> 1239 now = rq_clock_task(rq);
>> 1240 delta_exec = now - curr->se.exec_start;
>> 1241 if (unlikely((s64)delta_exec <= 0)) {
>> 1242 if (unlikely(dl_se->dl_yielded))
>> 1243 goto throttle;
>> 1244 return;
>> 1245 }
>> 1246
>> 1247 schedstat_set(curr->se.statistics.exec_max,
>> 1248 max(curr->se.statistics.exec_max, delta_exec));
>> 1249
>> 1250 curr->se.sum_exec_runtime += delta_exec;
>> 1251 account_group_exec_runtime(curr, delta_exec);
>> 1252
>> 1253 curr->se.exec_start = now;
>> 1254 cgroup_account_cputime(curr, delta_exec);
>> 1255
>> 1256 if (dl_entity_is_special(dl_se))
>> 1257 return;
1447 1258
1448 /* 1259 /*
1449 * For tasks that participate in GRUB 1260 * For tasks that participate in GRUB, we implement GRUB-PA: the
1450 * spare reclaimed bandwidth is used 1261 * spare reclaimed bandwidth is used to clock down frequency.
1451 * 1262 *
1452 * For the others, we still need to s 1263 * For the others, we still need to scale reservation parameters
1453 * according to current frequency and 1264 * according to current frequency and CPU maximum capacity.
1454 */ 1265 */
1455 if (unlikely(dl_se->flags & SCHED_FLA 1266 if (unlikely(dl_se->flags & SCHED_FLAG_RECLAIM)) {
1456 scaled_delta_exec = grub_recl !! 1267 scaled_delta_exec = grub_reclaim(delta_exec,
>> 1268 rq,
>> 1269 &curr->dl);
1457 } else { 1270 } else {
1458 int cpu = cpu_of(rq); <<
1459 unsigned long scale_freq = ar 1271 unsigned long scale_freq = arch_scale_freq_capacity(cpu);
1460 unsigned long scale_cpu = arc 1272 unsigned long scale_cpu = arch_scale_cpu_capacity(cpu);
1461 1273
1462 scaled_delta_exec = cap_scale 1274 scaled_delta_exec = cap_scale(delta_exec, scale_freq);
1463 scaled_delta_exec = cap_scale 1275 scaled_delta_exec = cap_scale(scaled_delta_exec, scale_cpu);
1464 } 1276 }
1465 1277
1466 return scaled_delta_exec; <<
1467 } <<
1468 <<
1469 static inline void <<
1470 update_stats_dequeue_dl(struct dl_rq *dl_rq, <<
1471 int flags); <<
1472 static void update_curr_dl_se(struct rq *rq, <<
1473 { <<
1474 s64 scaled_delta_exec; <<
1475 <<
1476 if (unlikely(delta_exec <= 0)) { <<
1477 if (unlikely(dl_se->dl_yielde <<
1478 goto throttle; <<
1479 return; <<
1480 } <<
1481 <<
1482 if (dl_server(dl_se) && dl_se->dl_thr <<
1483 return; <<
1484 <<
1485 if (dl_entity_is_special(dl_se)) <<
1486 return; <<
1487 <<
1488 scaled_delta_exec = dl_scaled_delta_e <<
1489 <<
1490 dl_se->runtime -= scaled_delta_exec; 1278 dl_se->runtime -= scaled_delta_exec;
1491 1279
1492 /* <<
1493 * The fair server can consume its ru <<
1494 * running as regular CFS). <<
1495 * <<
1496 * If the server consumes its entire <<
1497 * is not required for the current pe <<
1498 * starting a new period, pushing the <<
1499 */ <<
1500 if (dl_se->dl_defer && dl_se->dl_thro <<
1501 /* <<
1502 * If the server was previous <<
1503 * took place, it this point <<
1504 * was able to get runtime in <<
1505 * state. <<
1506 */ <<
1507 dl_se->dl_defer_running = 0; <<
1508 <<
1509 hrtimer_try_to_cancel(&dl_se- <<
1510 <<
1511 replenish_dl_new_period(dl_se <<
1512 <<
1513 /* <<
1514 * Not being able to start th <<
1515 * be started for whatever re <<
1516 * and queue right away. Othe <<
1517 * to what enqueue_dl_entity( <<
1518 */ <<
1519 WARN_ON_ONCE(!start_dl_timer( <<
1520 <<
1521 return; <<
1522 } <<
1523 <<
1524 throttle: 1280 throttle:
1525 if (dl_runtime_exceeded(dl_se) || dl_ 1281 if (dl_runtime_exceeded(dl_se) || dl_se->dl_yielded) {
1526 dl_se->dl_throttled = 1; 1282 dl_se->dl_throttled = 1;
1527 1283
1528 /* If requested, inform the u 1284 /* If requested, inform the user about runtime overruns. */
1529 if (dl_runtime_exceeded(dl_se 1285 if (dl_runtime_exceeded(dl_se) &&
1530 (dl_se->flags & SCHED_FLA 1286 (dl_se->flags & SCHED_FLAG_DL_OVERRUN))
1531 dl_se->dl_overrun = 1 1287 dl_se->dl_overrun = 1;
1532 1288
1533 dequeue_dl_entity(dl_se, 0); !! 1289 __dequeue_task_dl(rq, curr, 0);
1534 if (!dl_server(dl_se)) { !! 1290 if (unlikely(dl_se->dl_boosted || !start_dl_timer(curr)))
1535 update_stats_dequeue_ !! 1291 enqueue_task_dl(rq, curr, ENQUEUE_REPLENISH);
1536 dequeue_pushable_dl_t <<
1537 } <<
1538 1292
1539 if (unlikely(is_dl_boosted(dl !! 1293 if (!is_leftmost(curr, &rq->dl))
1540 if (dl_server(dl_se)) <<
1541 enqueue_dl_en <<
1542 else <<
1543 enqueue_task_ <<
1544 } <<
1545 <<
1546 if (!is_leftmost(dl_se, &rq-> <<
1547 resched_curr(rq); 1294 resched_curr(rq);
1548 } 1295 }
1549 1296
1550 /* 1297 /*
1551 * The fair server (sole dl_server) d <<
1552 * workload because it is running fai <<
1553 */ <<
1554 if (dl_se == &rq->fair_server) <<
1555 return; <<
1556 <<
1557 #ifdef CONFIG_RT_GROUP_SCHED <<
1558 /* <<
1559 * Because -- for now -- we share the 1298 * Because -- for now -- we share the rt bandwidth, we need to
1560 * account our runtime there too, oth 1299 * account our runtime there too, otherwise actual rt tasks
1561 * would be able to exceed the shared 1300 * would be able to exceed the shared quota.
1562 * 1301 *
1563 * Account to the root rt group for n 1302 * Account to the root rt group for now.
1564 * 1303 *
1565 * The solution we're working towards 1304 * The solution we're working towards is having the RT groups scheduled
1566 * using deadline servers -- however 1305 * using deadline servers -- however there's a few nasties to figure
1567 * out before that can happen. 1306 * out before that can happen.
1568 */ 1307 */
1569 if (rt_bandwidth_enabled()) { 1308 if (rt_bandwidth_enabled()) {
1570 struct rt_rq *rt_rq = &rq->rt 1309 struct rt_rq *rt_rq = &rq->rt;
1571 1310
1572 raw_spin_lock(&rt_rq->rt_runt 1311 raw_spin_lock(&rt_rq->rt_runtime_lock);
1573 /* 1312 /*
1574 * We'll let actual RT tasks 1313 * We'll let actual RT tasks worry about the overflow here, we
1575 * have our own CBS to keep u 1314 * have our own CBS to keep us inline; only account when RT
1576 * bandwidth is relevant. 1315 * bandwidth is relevant.
1577 */ 1316 */
1578 if (sched_rt_bandwidth_accoun 1317 if (sched_rt_bandwidth_account(rt_rq))
1579 rt_rq->rt_time += del 1318 rt_rq->rt_time += delta_exec;
1580 raw_spin_unlock(&rt_rq->rt_ru 1319 raw_spin_unlock(&rt_rq->rt_runtime_lock);
1581 } 1320 }
1582 #endif <<
1583 } <<
1584 <<
1585 /* <<
1586 * In the non-defer mode, the idle time is no <<
1587 * server provides a guarantee. <<
1588 * <<
1589 * If the dl_server is in defer mode, the idl <<
1590 * as time available for the fair server, avo <<
1591 * rt scheduler that did not consumed that ti <<
1592 */ <<
1593 void dl_server_update_idle_time(struct rq *rq <<
1594 { <<
1595 s64 delta_exec, scaled_delta_exec; <<
1596 <<
1597 if (!rq->fair_server.dl_defer) <<
1598 return; <<
1599 <<
1600 /* no need to discount more */ <<
1601 if (rq->fair_server.runtime < 0) <<
1602 return; <<
1603 <<
1604 delta_exec = rq_clock_task(rq) - p->s <<
1605 if (delta_exec < 0) <<
1606 return; <<
1607 <<
1608 scaled_delta_exec = dl_scaled_delta_e <<
1609 <<
1610 rq->fair_server.runtime -= scaled_del <<
1611 <<
1612 if (rq->fair_server.runtime < 0) { <<
1613 rq->fair_server.dl_defer_runn <<
1614 rq->fair_server.runtime = 0; <<
1615 } <<
1616 <<
1617 p->se.exec_start = rq_clock_task(rq); <<
1618 } <<
1619 <<
1620 void dl_server_update(struct sched_dl_entity <<
1621 { <<
1622 /* 0 runtime = fair server disabled * <<
1623 if (dl_se->dl_runtime) <<
1624 update_curr_dl_se(dl_se->rq, <<
1625 } <<
1626 <<
1627 void dl_server_start(struct sched_dl_entity * <<
1628 { <<
1629 struct rq *rq = dl_se->rq; <<
1630 <<
1631 /* <<
1632 * XXX: the apply do not work fine at <<
1633 * fair server because things are not <<
1634 * this before getting generic. <<
1635 */ <<
1636 if (!dl_server(dl_se)) { <<
1637 u64 runtime = 50 * NSEC_PER_ <<
1638 u64 period = 1000 * NSEC_PER_ <<
1639 <<
1640 dl_server_apply_params(dl_se, <<
1641 <<
1642 dl_se->dl_server = 1; <<
1643 dl_se->dl_defer = 1; <<
1644 setup_new_dl_entity(dl_se); <<
1645 } <<
1646 <<
1647 if (!dl_se->dl_runtime) <<
1648 return; <<
1649 <<
1650 enqueue_dl_entity(dl_se, ENQUEUE_WAKE <<
1651 if (!dl_task(dl_se->rq->curr) || dl_e <<
1652 resched_curr(dl_se->rq); <<
1653 } <<
1654 <<
1655 void dl_server_stop(struct sched_dl_entity *d <<
1656 { <<
1657 if (!dl_se->dl_runtime) <<
1658 return; <<
1659 <<
1660 dequeue_dl_entity(dl_se, DEQUEUE_SLEE <<
1661 hrtimer_try_to_cancel(&dl_se->dl_time <<
1662 dl_se->dl_defer_armed = 0; <<
1663 dl_se->dl_throttled = 0; <<
1664 } <<
1665 <<
1666 void dl_server_init(struct sched_dl_entity *d <<
1667 dl_server_has_tasks_f has <<
1668 dl_server_pick_f pick_tas <<
1669 { <<
1670 dl_se->rq = rq; <<
1671 dl_se->server_has_tasks = has_tasks; <<
1672 dl_se->server_pick_task = pick_task; <<
1673 } <<
1674 <<
1675 void __dl_server_attach_root(struct sched_dl_ <<
1676 { <<
1677 u64 new_bw = dl_se->dl_bw; <<
1678 int cpu = cpu_of(rq); <<
1679 struct dl_bw *dl_b; <<
1680 <<
1681 dl_b = dl_bw_of(cpu_of(rq)); <<
1682 guard(raw_spinlock)(&dl_b->lock); <<
1683 <<
1684 if (!dl_bw_cpus(cpu)) <<
1685 return; <<
1686 <<
1687 __dl_add(dl_b, new_bw, dl_bw_cpus(cpu <<
1688 } <<
1689 <<
1690 int dl_server_apply_params(struct sched_dl_en <<
1691 { <<
1692 u64 old_bw = init ? 0 : to_ratio(dl_s <<
1693 u64 new_bw = to_ratio(period, runtime <<
1694 struct rq *rq = dl_se->rq; <<
1695 int cpu = cpu_of(rq); <<
1696 struct dl_bw *dl_b; <<
1697 unsigned long cap; <<
1698 int retval = 0; <<
1699 int cpus; <<
1700 <<
1701 dl_b = dl_bw_of(cpu); <<
1702 guard(raw_spinlock)(&dl_b->lock); <<
1703 <<
1704 cpus = dl_bw_cpus(cpu); <<
1705 cap = dl_bw_capacity(cpu); <<
1706 <<
1707 if (__dl_overflow(dl_b, cap, old_bw, <<
1708 return -EBUSY; <<
1709 <<
1710 if (init) { <<
1711 __add_rq_bw(new_bw, &rq->dl); <<
1712 __dl_add(dl_b, new_bw, cpus); <<
1713 } else { <<
1714 __dl_sub(dl_b, dl_se->dl_bw, <<
1715 __dl_add(dl_b, new_bw, cpus); <<
1716 <<
1717 dl_rq_change_utilization(rq, <<
1718 } <<
1719 <<
1720 dl_se->dl_runtime = runtime; <<
1721 dl_se->dl_deadline = period; <<
1722 dl_se->dl_period = period; <<
1723 <<
1724 dl_se->runtime = 0; <<
1725 dl_se->deadline = 0; <<
1726 <<
1727 dl_se->dl_bw = to_ratio(dl_se->dl_per <<
1728 dl_se->dl_density = to_ratio(dl_se->d <<
1729 <<
1730 return retval; <<
1731 } <<
1732 <<
1733 /* <<
1734 * Update the current task's runtime statisti <<
1735 * a -deadline task and has not been removed <<
1736 */ <<
1737 static void update_curr_dl(struct rq *rq) <<
1738 { <<
1739 struct task_struct *curr = rq->curr; <<
1740 struct sched_dl_entity *dl_se = &curr <<
1741 s64 delta_exec; <<
1742 <<
1743 if (!dl_task(curr) || !on_dl_rq(dl_se <<
1744 return; <<
1745 <<
1746 /* <<
1747 * Consumed budget is computed consid <<
1748 * observed by schedulable tasks (exc <<
1749 * in hardirq context, etc.). Deadlin <<
1750 * computed using hard walltime. This <<
1751 * natural solution, but the full ram <<
1752 * approach need further study. <<
1753 */ <<
1754 delta_exec = update_curr_common(rq); <<
1755 update_curr_dl_se(rq, dl_se, delta_ex <<
1756 } 1321 }
1757 1322
1758 static enum hrtimer_restart inactive_task_tim 1323 static enum hrtimer_restart inactive_task_timer(struct hrtimer *timer)
1759 { 1324 {
1760 struct sched_dl_entity *dl_se = conta 1325 struct sched_dl_entity *dl_se = container_of(timer,
1761 1326 struct sched_dl_entity,
1762 1327 inactive_timer);
1763 struct task_struct *p = NULL; !! 1328 struct task_struct *p = dl_task_of(dl_se);
1764 struct rq_flags rf; 1329 struct rq_flags rf;
1765 struct rq *rq; 1330 struct rq *rq;
1766 1331
1767 if (!dl_server(dl_se)) { !! 1332 rq = task_rq_lock(p, &rf);
1768 p = dl_task_of(dl_se); <<
1769 rq = task_rq_lock(p, &rf); <<
1770 } else { <<
1771 rq = dl_se->rq; <<
1772 rq_lock(rq, &rf); <<
1773 } <<
1774 1333
1775 sched_clock_tick(); 1334 sched_clock_tick();
1776 update_rq_clock(rq); 1335 update_rq_clock(rq);
1777 1336
1778 if (dl_server(dl_se)) !! 1337 if (!dl_task(p) || p->state == TASK_DEAD) {
1779 goto no_task; <<
1780 <<
1781 if (!dl_task(p) || READ_ONCE(p->__sta <<
1782 struct dl_bw *dl_b = dl_bw_of 1338 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1783 1339
1784 if (READ_ONCE(p->__state) == !! 1340 if (p->state == TASK_DEAD && dl_se->dl_non_contending) {
1785 sub_running_bw(&p->dl 1341 sub_running_bw(&p->dl, dl_rq_of_se(&p->dl));
1786 sub_rq_bw(&p->dl, dl_ 1342 sub_rq_bw(&p->dl, dl_rq_of_se(&p->dl));
1787 dl_se->dl_non_contend 1343 dl_se->dl_non_contending = 0;
1788 } 1344 }
1789 1345
1790 raw_spin_lock(&dl_b->lock); 1346 raw_spin_lock(&dl_b->lock);
1791 __dl_sub(dl_b, p->dl.dl_bw, d 1347 __dl_sub(dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
1792 raw_spin_unlock(&dl_b->lock); 1348 raw_spin_unlock(&dl_b->lock);
1793 __dl_clear_params(dl_se); !! 1349 __dl_clear_params(p);
1794 1350
1795 goto unlock; 1351 goto unlock;
1796 } 1352 }
1797 <<
1798 no_task: <<
1799 if (dl_se->dl_non_contending == 0) 1353 if (dl_se->dl_non_contending == 0)
1800 goto unlock; 1354 goto unlock;
1801 1355
1802 sub_running_bw(dl_se, &rq->dl); 1356 sub_running_bw(dl_se, &rq->dl);
1803 dl_se->dl_non_contending = 0; 1357 dl_se->dl_non_contending = 0;
1804 unlock: 1358 unlock:
1805 !! 1359 task_rq_unlock(rq, p, &rf);
1806 if (!dl_server(dl_se)) { !! 1360 put_task_struct(p);
1807 task_rq_unlock(rq, p, &rf); <<
1808 put_task_struct(p); <<
1809 } else { <<
1810 rq_unlock(rq, &rf); <<
1811 } <<
1812 1361
1813 return HRTIMER_NORESTART; 1362 return HRTIMER_NORESTART;
1814 } 1363 }
1815 1364
1816 static void init_dl_inactive_task_timer(struc !! 1365 void init_dl_inactive_task_timer(struct sched_dl_entity *dl_se)
1817 { 1366 {
1818 struct hrtimer *timer = &dl_se->inact 1367 struct hrtimer *timer = &dl_se->inactive_timer;
1819 1368
1820 hrtimer_init(timer, CLOCK_MONOTONIC, 1369 hrtimer_init(timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
1821 timer->function = inactive_task_timer 1370 timer->function = inactive_task_timer;
1822 } 1371 }
1823 1372
1824 #define __node_2_dle(node) \ <<
1825 rb_entry((node), struct sched_dl_enti <<
1826 <<
1827 #ifdef CONFIG_SMP 1373 #ifdef CONFIG_SMP
1828 1374
1829 static void inc_dl_deadline(struct dl_rq *dl_ 1375 static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1830 { 1376 {
1831 struct rq *rq = rq_of_dl_rq(dl_rq); 1377 struct rq *rq = rq_of_dl_rq(dl_rq);
1832 1378
1833 if (dl_rq->earliest_dl.curr == 0 || 1379 if (dl_rq->earliest_dl.curr == 0 ||
1834 dl_time_before(deadline, dl_rq->e 1380 dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
1835 if (dl_rq->earliest_dl.curr = <<
1836 cpupri_set(&rq->rd->c <<
1837 dl_rq->earliest_dl.curr = dea 1381 dl_rq->earliest_dl.curr = deadline;
1838 cpudl_set(&rq->rd->cpudl, rq- 1382 cpudl_set(&rq->rd->cpudl, rq->cpu, deadline);
1839 } 1383 }
1840 } 1384 }
1841 1385
1842 static void dec_dl_deadline(struct dl_rq *dl_ 1386 static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
1843 { 1387 {
1844 struct rq *rq = rq_of_dl_rq(dl_rq); 1388 struct rq *rq = rq_of_dl_rq(dl_rq);
1845 1389
1846 /* 1390 /*
1847 * Since we may have removed our earl 1391 * Since we may have removed our earliest (and/or next earliest)
1848 * task we must recompute them. 1392 * task we must recompute them.
1849 */ 1393 */
1850 if (!dl_rq->dl_nr_running) { 1394 if (!dl_rq->dl_nr_running) {
1851 dl_rq->earliest_dl.curr = 0; 1395 dl_rq->earliest_dl.curr = 0;
1852 dl_rq->earliest_dl.next = 0; 1396 dl_rq->earliest_dl.next = 0;
1853 cpudl_clear(&rq->rd->cpudl, r 1397 cpudl_clear(&rq->rd->cpudl, rq->cpu);
1854 cpupri_set(&rq->rd->cpupri, r <<
1855 } else { 1398 } else {
1856 struct rb_node *leftmost = rb !! 1399 struct rb_node *leftmost = dl_rq->root.rb_leftmost;
1857 struct sched_dl_entity *entry !! 1400 struct sched_dl_entity *entry;
1858 1401
>> 1402 entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
1859 dl_rq->earliest_dl.curr = ent 1403 dl_rq->earliest_dl.curr = entry->deadline;
1860 cpudl_set(&rq->rd->cpudl, rq- 1404 cpudl_set(&rq->rd->cpudl, rq->cpu, entry->deadline);
1861 } 1405 }
1862 } 1406 }
1863 1407
1864 #else 1408 #else
1865 1409
1866 static inline void inc_dl_deadline(struct dl_ 1410 static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1867 static inline void dec_dl_deadline(struct dl_ 1411 static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
1868 1412
1869 #endif /* CONFIG_SMP */ 1413 #endif /* CONFIG_SMP */
1870 1414
1871 static inline 1415 static inline
1872 void inc_dl_tasks(struct sched_dl_entity *dl_ 1416 void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1873 { 1417 {
>> 1418 int prio = dl_task_of(dl_se)->prio;
1874 u64 deadline = dl_se->deadline; 1419 u64 deadline = dl_se->deadline;
1875 1420
>> 1421 WARN_ON(!dl_prio(prio));
1876 dl_rq->dl_nr_running++; 1422 dl_rq->dl_nr_running++;
1877 add_nr_running(rq_of_dl_rq(dl_rq), 1) 1423 add_nr_running(rq_of_dl_rq(dl_rq), 1);
1878 1424
1879 inc_dl_deadline(dl_rq, deadline); 1425 inc_dl_deadline(dl_rq, deadline);
>> 1426 inc_dl_migration(dl_se, dl_rq);
1880 } 1427 }
1881 1428
1882 static inline 1429 static inline
1883 void dec_dl_tasks(struct sched_dl_entity *dl_ 1430 void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
1884 { 1431 {
>> 1432 int prio = dl_task_of(dl_se)->prio;
>> 1433
>> 1434 WARN_ON(!dl_prio(prio));
1885 WARN_ON(!dl_rq->dl_nr_running); 1435 WARN_ON(!dl_rq->dl_nr_running);
1886 dl_rq->dl_nr_running--; 1436 dl_rq->dl_nr_running--;
1887 sub_nr_running(rq_of_dl_rq(dl_rq), 1) 1437 sub_nr_running(rq_of_dl_rq(dl_rq), 1);
1888 1438
1889 dec_dl_deadline(dl_rq, dl_se->deadlin 1439 dec_dl_deadline(dl_rq, dl_se->deadline);
1890 } !! 1440 dec_dl_migration(dl_se, dl_rq);
1891 <<
1892 static inline bool __dl_less(struct rb_node * <<
1893 { <<
1894 return dl_time_before(__node_2_dle(a) <<
1895 } <<
1896 <<
1897 static __always_inline struct sched_statistic <<
1898 __schedstats_from_dl_se(struct sched_dl_entit <<
1899 { <<
1900 if (!schedstat_enabled()) <<
1901 return NULL; <<
1902 <<
1903 if (dl_server(dl_se)) <<
1904 return NULL; <<
1905 <<
1906 return &dl_task_of(dl_se)->stats; <<
1907 } <<
1908 <<
1909 static inline void <<
1910 update_stats_wait_start_dl(struct dl_rq *dl_r <<
1911 { <<
1912 struct sched_statistics *stats = __sc <<
1913 if (stats) <<
1914 __update_stats_wait_start(rq_ <<
1915 } <<
1916 <<
1917 static inline void <<
1918 update_stats_wait_end_dl(struct dl_rq *dl_rq, <<
1919 { <<
1920 struct sched_statistics *stats = __sc <<
1921 if (stats) <<
1922 __update_stats_wait_end(rq_of <<
1923 } <<
1924 <<
1925 static inline void <<
1926 update_stats_enqueue_sleeper_dl(struct dl_rq <<
1927 { <<
1928 struct sched_statistics *stats = __sc <<
1929 if (stats) <<
1930 __update_stats_enqueue_sleepe <<
1931 } <<
1932 <<
1933 static inline void <<
1934 update_stats_enqueue_dl(struct dl_rq *dl_rq, <<
1935 int flags) <<
1936 { <<
1937 if (!schedstat_enabled()) <<
1938 return; <<
1939 <<
1940 if (flags & ENQUEUE_WAKEUP) <<
1941 update_stats_enqueue_sleeper_ <<
1942 } <<
1943 <<
1944 static inline void <<
1945 update_stats_dequeue_dl(struct dl_rq *dl_rq, <<
1946 int flags) <<
1947 { <<
1948 struct task_struct *p = dl_task_of(dl <<
1949 <<
1950 if (!schedstat_enabled()) <<
1951 return; <<
1952 <<
1953 if ((flags & DEQUEUE_SLEEP)) { <<
1954 unsigned int state; <<
1955 <<
1956 state = READ_ONCE(p->__state) <<
1957 if (state & TASK_INTERRUPTIBL <<
1958 __schedstat_set(p->st <<
1959 rq_cl <<
1960 <<
1961 if (state & TASK_UNINTERRUPTI <<
1962 __schedstat_set(p->st <<
1963 rq_cl <<
1964 } <<
1965 } 1441 }
1966 1442
1967 static void __enqueue_dl_entity(struct sched_ 1443 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
1968 { 1444 {
1969 struct dl_rq *dl_rq = dl_rq_of_se(dl_ 1445 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
>> 1446 struct rb_node **link = &dl_rq->root.rb_root.rb_node;
>> 1447 struct rb_node *parent = NULL;
>> 1448 struct sched_dl_entity *entry;
>> 1449 int leftmost = 1;
>> 1450
>> 1451 BUG_ON(!RB_EMPTY_NODE(&dl_se->rb_node));
>> 1452
>> 1453 while (*link) {
>> 1454 parent = *link;
>> 1455 entry = rb_entry(parent, struct sched_dl_entity, rb_node);
>> 1456 if (dl_time_before(dl_se->deadline, entry->deadline))
>> 1457 link = &parent->rb_left;
>> 1458 else {
>> 1459 link = &parent->rb_right;
>> 1460 leftmost = 0;
>> 1461 }
>> 1462 }
1970 1463
1971 WARN_ON_ONCE(!RB_EMPTY_NODE(&dl_se->r !! 1464 rb_link_node(&dl_se->rb_node, parent, link);
1972 !! 1465 rb_insert_color_cached(&dl_se->rb_node, &dl_rq->root, leftmost);
1973 rb_add_cached(&dl_se->rb_node, &dl_rq <<
1974 1466
1975 inc_dl_tasks(dl_se, dl_rq); 1467 inc_dl_tasks(dl_se, dl_rq);
1976 } 1468 }
1977 1469
1978 static void __dequeue_dl_entity(struct sched_ 1470 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
1979 { 1471 {
1980 struct dl_rq *dl_rq = dl_rq_of_se(dl_ 1472 struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
1981 1473
1982 if (RB_EMPTY_NODE(&dl_se->rb_node)) 1474 if (RB_EMPTY_NODE(&dl_se->rb_node))
1983 return; 1475 return;
1984 1476
1985 rb_erase_cached(&dl_se->rb_node, &dl_ 1477 rb_erase_cached(&dl_se->rb_node, &dl_rq->root);
1986 <<
1987 RB_CLEAR_NODE(&dl_se->rb_node); 1478 RB_CLEAR_NODE(&dl_se->rb_node);
1988 1479
1989 dec_dl_tasks(dl_se, dl_rq); 1480 dec_dl_tasks(dl_se, dl_rq);
1990 } 1481 }
1991 1482
1992 static void 1483 static void
1993 enqueue_dl_entity(struct sched_dl_entity *dl_ !! 1484 enqueue_dl_entity(struct sched_dl_entity *dl_se,
>> 1485 struct sched_dl_entity *pi_se, int flags)
>> 1486 {
>> 1487 BUG_ON(on_dl_rq(dl_se));
>> 1488
>> 1489 /*
>> 1490 * If this is a wakeup or a new instance, the scheduling
>> 1491 * parameters of the task might need updating. Otherwise,
>> 1492 * we want a replenishment of its runtime.
>> 1493 */
>> 1494 if (flags & ENQUEUE_WAKEUP) {
>> 1495 task_contending(dl_se, flags);
>> 1496 update_dl_entity(dl_se, pi_se);
>> 1497 } else if (flags & ENQUEUE_REPLENISH) {
>> 1498 replenish_dl_entity(dl_se, pi_se);
>> 1499 } else if ((flags & ENQUEUE_RESTORE) &&
>> 1500 dl_time_before(dl_se->deadline,
>> 1501 rq_clock(rq_of_dl_rq(dl_rq_of_se(dl_se))))) {
>> 1502 setup_new_dl_entity(dl_se);
>> 1503 }
>> 1504
>> 1505 __enqueue_dl_entity(dl_se);
>> 1506 }
>> 1507
>> 1508 static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
>> 1509 {
>> 1510 __dequeue_dl_entity(dl_se);
>> 1511 }
>> 1512
>> 1513 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
1994 { 1514 {
1995 WARN_ON_ONCE(on_dl_rq(dl_se)); !! 1515 struct task_struct *pi_task = rt_mutex_get_top_task(p);
>> 1516 struct sched_dl_entity *pi_se = &p->dl;
1996 1517
1997 update_stats_enqueue_dl(dl_rq_of_se(d !! 1518 /*
>> 1519 * Use the scheduling parameters of the top pi-waiter task if:
>> 1520 * - we have a top pi-waiter which is a SCHED_DEADLINE task AND
>> 1521 * - our dl_boosted is set (i.e. the pi-waiter's (absolute) deadline is
>> 1522 * smaller than our deadline OR we are a !SCHED_DEADLINE task getting
>> 1523 * boosted due to a SCHED_DEADLINE pi-waiter).
>> 1524 * Otherwise we keep our runtime and deadline.
>> 1525 */
>> 1526 if (pi_task && dl_prio(pi_task->normal_prio) && p->dl.dl_boosted) {
>> 1527 pi_se = &pi_task->dl;
>> 1528 } else if (!dl_prio(p->normal_prio)) {
>> 1529 /*
>> 1530 * Special case in which we have a !SCHED_DEADLINE task
>> 1531 * that is going to be deboosted, but exceeds its
>> 1532 * runtime while doing so. No point in replenishing
>> 1533 * it, as it's going to return back to its original
>> 1534 * scheduling class after this.
>> 1535 */
>> 1536 BUG_ON(!p->dl.dl_boosted || flags != ENQUEUE_REPLENISH);
>> 1537 return;
>> 1538 }
1998 1539
1999 /* 1540 /*
2000 * Check if a constrained deadline ta 1541 * Check if a constrained deadline task was activated
2001 * after the deadline but before the 1542 * after the deadline but before the next period.
2002 * If that is the case, the task will 1543 * If that is the case, the task will be throttled and
2003 * the replenishment timer will be se 1544 * the replenishment timer will be set to the next period.
2004 */ 1545 */
2005 if (!dl_se->dl_throttled && !dl_is_im !! 1546 if (!p->dl.dl_throttled && !dl_is_implicit(&p->dl))
2006 dl_check_constrained_dl(dl_se !! 1547 dl_check_constrained_dl(&p->dl);
2007 1548
2008 if (flags & (ENQUEUE_RESTORE|ENQUEUE_ !! 1549 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & ENQUEUE_RESTORE) {
2009 struct dl_rq *dl_rq = dl_rq_o !! 1550 add_rq_bw(&p->dl, &rq->dl);
2010 !! 1551 add_running_bw(&p->dl, &rq->dl);
2011 add_rq_bw(dl_se, dl_rq); <<
2012 add_running_bw(dl_se, dl_rq); <<
2013 } 1552 }
2014 1553
2015 /* 1554 /*
2016 * If p is throttled, we do not enque 1555 * If p is throttled, we do not enqueue it. In fact, if it exhausted
2017 * its budget it needs a replenishmen 1556 * its budget it needs a replenishment and, since it now is on
2018 * its rq, the bandwidth timer callba 1557 * its rq, the bandwidth timer callback (which clearly has not
2019 * run yet) will take care of this. 1558 * run yet) will take care of this.
2020 * However, the active utilization do 1559 * However, the active utilization does not depend on the fact
2021 * that the task is on the runqueue o 1560 * that the task is on the runqueue or not (but depends on the
2022 * task's state - in GRUB parlance, " 1561 * task's state - in GRUB parlance, "inactive" vs "active contending").
2023 * In other words, even if a task is 1562 * In other words, even if a task is throttled its utilization must
2024 * be counted in the active utilizati 1563 * be counted in the active utilization; hence, we need to call
2025 * add_running_bw(). 1564 * add_running_bw().
2026 */ 1565 */
2027 if (!dl_se->dl_defer && dl_se->dl_thr !! 1566 if (p->dl.dl_throttled && !(flags & ENQUEUE_REPLENISH)) {
2028 if (flags & ENQUEUE_WAKEUP) 1567 if (flags & ENQUEUE_WAKEUP)
2029 task_contending(dl_se !! 1568 task_contending(&p->dl, flags);
2030 1569
2031 return; 1570 return;
2032 } 1571 }
2033 1572
2034 /* !! 1573 enqueue_dl_entity(&p->dl, pi_se, flags);
2035 * If this is a wakeup or a new insta <<
2036 * parameters of the task might need <<
2037 * we want a replenishment of its run <<
2038 */ <<
2039 if (flags & ENQUEUE_WAKEUP) { <<
2040 task_contending(dl_se, flags) <<
2041 update_dl_entity(dl_se); <<
2042 } else if (flags & ENQUEUE_REPLENISH) <<
2043 replenish_dl_entity(dl_se); <<
2044 } else if ((flags & ENQUEUE_RESTORE) <<
2045 dl_time_before(dl_se->dead <<
2046 setup_new_dl_entity(dl_se); <<
2047 } <<
2048 <<
2049 /* <<
2050 * If the reservation is still thrott <<
2051 * deferred task and still got to wai <<
2052 */ <<
2053 if (dl_se->dl_throttled && start_dl_t <<
2054 return; <<
2055 1574
2056 /* !! 1575 if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
2057 * We're about to enqueue, make sure !! 1576 enqueue_pushable_dl_task(rq, p);
2058 * In case the timer was not started, <<
2059 * has passed, mark as not throttled <<
2060 * Also cancel earlier timers, since <<
2061 */ <<
2062 if (dl_se->dl_throttled) { <<
2063 hrtimer_try_to_cancel(&dl_se- <<
2064 dl_se->dl_defer_armed = 0; <<
2065 dl_se->dl_throttled = 0; <<
2066 } <<
2067 <<
2068 __enqueue_dl_entity(dl_se); <<
2069 } 1577 }
2070 1578
2071 static void dequeue_dl_entity(struct sched_dl !! 1579 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2072 { 1580 {
2073 __dequeue_dl_entity(dl_se); !! 1581 dequeue_dl_entity(&p->dl);
>> 1582 dequeue_pushable_dl_task(rq, p);
>> 1583 }
2074 1584
2075 if (flags & (DEQUEUE_SAVE|DEQUEUE_MIG !! 1585 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
2076 struct dl_rq *dl_rq = dl_rq_o !! 1586 {
>> 1587 update_curr_dl(rq);
>> 1588 __dequeue_task_dl(rq, p, flags);
2077 1589
2078 sub_running_bw(dl_se, dl_rq); !! 1590 if (p->on_rq == TASK_ON_RQ_MIGRATING || flags & DEQUEUE_SAVE) {
2079 sub_rq_bw(dl_se, dl_rq); !! 1591 sub_running_bw(&p->dl, &rq->dl);
>> 1592 sub_rq_bw(&p->dl, &rq->dl);
2080 } 1593 }
2081 1594
2082 /* 1595 /*
2083 * This check allows to start the ina 1596 * This check allows to start the inactive timer (or to immediately
2084 * decrease the active utilization, i 1597 * decrease the active utilization, if needed) in two cases:
2085 * when the task blocks and when it i 1598 * when the task blocks and when it is terminating
2086 * (p->state == TASK_DEAD). We can ha 1599 * (p->state == TASK_DEAD). We can handle the two cases in the same
2087 * way, because from GRUB's point of 1600 * way, because from GRUB's point of view the same thing is happening
2088 * (the task moves from "active conte 1601 * (the task moves from "active contending" to "active non contending"
2089 * or "inactive") 1602 * or "inactive")
2090 */ 1603 */
2091 if (flags & DEQUEUE_SLEEP) 1604 if (flags & DEQUEUE_SLEEP)
2092 task_non_contending(dl_se); !! 1605 task_non_contending(p);
2093 } <<
2094 <<
2095 static void enqueue_task_dl(struct rq *rq, st <<
2096 { <<
2097 if (is_dl_boosted(&p->dl)) { <<
2098 /* <<
2099 * Because of delays in the d <<
2100 * thread's runtime, it might <<
2101 * goes to sleep in a rt mute <<
2102 * a consequence, the thread <<
2103 * <<
2104 * While waiting for the mute <<
2105 * boosted via PI, resulting <<
2106 * and boosted at the same ti <<
2107 * <<
2108 * In this case, the boost ov <<
2109 */ <<
2110 if (p->dl.dl_throttled) { <<
2111 /* <<
2112 * The replenish time <<
2113 * problem if it fire <<
2114 * are ignored in dl_ <<
2115 * <<
2116 * If the timer callb <<
2117 * it will eventually <<
2118 */ <<
2119 if (hrtimer_try_to_ca <<
2120 !dl_server(&p->dl <<
2121 put_task_stru <<
2122 p->dl.dl_throttled = <<
2123 } <<
2124 } else if (!dl_prio(p->normal_prio)) <<
2125 /* <<
2126 * Special case in which we h <<
2127 * to be deboosted, but excee <<
2128 * replenishing it, as it's g <<
2129 * scheduling class after thi <<
2130 * clear the flag, otherwise <<
2131 * being boosted again with n <<
2132 * the throttle. <<
2133 */ <<
2134 p->dl.dl_throttled = 0; <<
2135 if (!(flags & ENQUEUE_REPLENI <<
2136 printk_deferred_once( <<
2137 <<
2138 <<
2139 return; <<
2140 } <<
2141 <<
2142 check_schedstat_required(); <<
2143 update_stats_wait_start_dl(dl_rq_of_s <<
2144 <<
2145 if (p->on_rq == TASK_ON_RQ_MIGRATING) <<
2146 flags |= ENQUEUE_MIGRATING; <<
2147 <<
2148 enqueue_dl_entity(&p->dl, flags); <<
2149 <<
2150 if (dl_server(&p->dl)) <<
2151 return; <<
2152 <<
2153 if (!task_current(rq, p) && !p->dl.dl <<
2154 enqueue_pushable_dl_task(rq, <<
2155 } <<
2156 <<
2157 static bool dequeue_task_dl(struct rq *rq, st <<
2158 { <<
2159 update_curr_dl(rq); <<
2160 <<
2161 if (p->on_rq == TASK_ON_RQ_MIGRATING) <<
2162 flags |= DEQUEUE_MIGRATING; <<
2163 <<
2164 dequeue_dl_entity(&p->dl, flags); <<
2165 if (!p->dl.dl_throttled && !dl_server <<
2166 dequeue_pushable_dl_task(rq, <<
2167 <<
2168 return true; <<
2169 } 1606 }
2170 1607
2171 /* 1608 /*
2172 * Yield task semantic for -deadline tasks is 1609 * Yield task semantic for -deadline tasks is:
2173 * 1610 *
2174 * get off from the CPU until our next inst 1611 * get off from the CPU until our next instance, with
2175 * a new runtime. This is of little use now 1612 * a new runtime. This is of little use now, since we
2176 * don't have a bandwidth reclaiming mechan 1613 * don't have a bandwidth reclaiming mechanism. Anyway,
2177 * bandwidth reclaiming is planned for the 1614 * bandwidth reclaiming is planned for the future, and
2178 * yield_task_dl will indicate that some sp 1615 * yield_task_dl will indicate that some spare budget
2179 * is available for other task instances to 1616 * is available for other task instances to use it.
2180 */ 1617 */
2181 static void yield_task_dl(struct rq *rq) 1618 static void yield_task_dl(struct rq *rq)
2182 { 1619 {
2183 /* 1620 /*
2184 * We make the task go to sleep until 1621 * We make the task go to sleep until its current deadline by
2185 * forcing its runtime to zero. This 1622 * forcing its runtime to zero. This way, update_curr_dl() stops
2186 * it and the bandwidth timer will wa 1623 * it and the bandwidth timer will wake it up and will give it
2187 * new scheduling parameters (thanks 1624 * new scheduling parameters (thanks to dl_yielded=1).
2188 */ 1625 */
2189 rq->curr->dl.dl_yielded = 1; 1626 rq->curr->dl.dl_yielded = 1;
2190 1627
2191 update_rq_clock(rq); 1628 update_rq_clock(rq);
2192 update_curr_dl(rq); 1629 update_curr_dl(rq);
2193 /* 1630 /*
2194 * Tell update_rq_clock() that we've 1631 * Tell update_rq_clock() that we've just updated,
2195 * so we don't do microscopic update 1632 * so we don't do microscopic update in schedule()
2196 * and double the fastpath cost. 1633 * and double the fastpath cost.
2197 */ 1634 */
2198 rq_clock_skip_update(rq); 1635 rq_clock_skip_update(rq);
2199 } 1636 }
2200 1637
2201 #ifdef CONFIG_SMP 1638 #ifdef CONFIG_SMP
2202 1639
2203 static inline bool dl_task_is_earliest_deadli <<
2204 <<
2205 { <<
2206 return (!rq->dl.dl_nr_running || <<
2207 dl_time_before(p->dl.deadline <<
2208 rq->dl.earlies <<
2209 } <<
2210 <<
2211 static int find_later_rq(struct task_struct * 1640 static int find_later_rq(struct task_struct *task);
2212 1641
2213 static int 1642 static int
2214 select_task_rq_dl(struct task_struct *p, int !! 1643 select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
2215 { 1644 {
2216 struct task_struct *curr; 1645 struct task_struct *curr;
2217 bool select_rq; 1646 bool select_rq;
2218 struct rq *rq; 1647 struct rq *rq;
2219 1648
2220 if (!(flags & WF_TTWU)) !! 1649 if (sd_flag != SD_BALANCE_WAKE)
2221 goto out; 1650 goto out;
2222 1651
2223 rq = cpu_rq(cpu); 1652 rq = cpu_rq(cpu);
2224 1653
2225 rcu_read_lock(); 1654 rcu_read_lock();
2226 curr = READ_ONCE(rq->curr); /* unlock 1655 curr = READ_ONCE(rq->curr); /* unlocked access */
2227 1656
2228 /* 1657 /*
2229 * If we are dealing with a -deadline 1658 * If we are dealing with a -deadline task, we must
2230 * decide where to wake it up. 1659 * decide where to wake it up.
2231 * If it has a later deadline and the 1660 * If it has a later deadline and the current task
2232 * on this rq can't move (provided th 1661 * on this rq can't move (provided the waking task
2233 * can!) we prefer to send it somewhe 1662 * can!) we prefer to send it somewhere else. On the
2234 * other hand, if it has a shorter de 1663 * other hand, if it has a shorter deadline, we
2235 * try to make it stay here, it might 1664 * try to make it stay here, it might be important.
2236 */ 1665 */
2237 select_rq = unlikely(dl_task(curr)) & 1666 select_rq = unlikely(dl_task(curr)) &&
2238 (curr->nr_cpus_allowed < 1667 (curr->nr_cpus_allowed < 2 ||
2239 !dl_entity_preempt(&p->d 1668 !dl_entity_preempt(&p->dl, &curr->dl)) &&
2240 p->nr_cpus_allowed > 1; 1669 p->nr_cpus_allowed > 1;
2241 1670
2242 /* 1671 /*
2243 * Take the capacity of the CPU into 1672 * Take the capacity of the CPU into account to
2244 * ensure it fits the requirement of 1673 * ensure it fits the requirement of the task.
2245 */ 1674 */
2246 if (sched_asym_cpucap_active()) !! 1675 if (static_branch_unlikely(&sched_asym_cpucapacity))
2247 select_rq |= !dl_task_fits_ca 1676 select_rq |= !dl_task_fits_capacity(p, cpu);
2248 1677
2249 if (select_rq) { 1678 if (select_rq) {
2250 int target = find_later_rq(p) 1679 int target = find_later_rq(p);
2251 1680
2252 if (target != -1 && 1681 if (target != -1 &&
2253 dl_task_is_earliest_deadl !! 1682 (dl_time_before(p->dl.deadline,
>> 1683 cpu_rq(target)->dl.earliest_dl.curr) ||
>> 1684 (cpu_rq(target)->dl.dl_nr_running == 0)))
2254 cpu = target; 1685 cpu = target;
2255 } 1686 }
2256 rcu_read_unlock(); 1687 rcu_read_unlock();
2257 1688
2258 out: 1689 out:
2259 return cpu; 1690 return cpu;
2260 } 1691 }
2261 1692
2262 static void migrate_task_rq_dl(struct task_st 1693 static void migrate_task_rq_dl(struct task_struct *p, int new_cpu __maybe_unused)
2263 { 1694 {
2264 struct rq_flags rf; <<
2265 struct rq *rq; 1695 struct rq *rq;
2266 1696
2267 if (READ_ONCE(p->__state) != TASK_WAK !! 1697 if (p->state != TASK_WAKING)
2268 return; 1698 return;
2269 1699
2270 rq = task_rq(p); 1700 rq = task_rq(p);
2271 /* 1701 /*
2272 * Since p->state == TASK_WAKING, set 1702 * Since p->state == TASK_WAKING, set_task_cpu() has been called
2273 * from try_to_wake_up(). Hence, p->p 1703 * from try_to_wake_up(). Hence, p->pi_lock is locked, but
2274 * rq->lock is not... So, lock it 1704 * rq->lock is not... So, lock it
2275 */ 1705 */
2276 rq_lock(rq, &rf); !! 1706 raw_spin_lock(&rq->lock);
2277 if (p->dl.dl_non_contending) { 1707 if (p->dl.dl_non_contending) {
2278 update_rq_clock(rq); <<
2279 sub_running_bw(&p->dl, &rq->d 1708 sub_running_bw(&p->dl, &rq->dl);
2280 p->dl.dl_non_contending = 0; 1709 p->dl.dl_non_contending = 0;
2281 /* 1710 /*
2282 * If the timer handler is cu 1711 * If the timer handler is currently running and the
2283 * timer cannot be canceled, !! 1712 * timer cannot be cancelled, inactive_task_timer()
2284 * will see that dl_not_conte 1713 * will see that dl_not_contending is not set, and
2285 * will not touch the rq's ac 1714 * will not touch the rq's active utilization,
2286 * so we are still safe. 1715 * so we are still safe.
2287 */ 1716 */
2288 if (hrtimer_try_to_cancel(&p- 1717 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
2289 put_task_struct(p); 1718 put_task_struct(p);
2290 } 1719 }
2291 sub_rq_bw(&p->dl, &rq->dl); 1720 sub_rq_bw(&p->dl, &rq->dl);
2292 rq_unlock(rq, &rf); !! 1721 raw_spin_unlock(&rq->lock);
2293 } 1722 }
2294 1723
2295 static void check_preempt_equal_dl(struct rq 1724 static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
2296 { 1725 {
2297 /* 1726 /*
2298 * Current can't be migrated, useless 1727 * Current can't be migrated, useless to reschedule,
2299 * let's hope p can move out. 1728 * let's hope p can move out.
2300 */ 1729 */
2301 if (rq->curr->nr_cpus_allowed == 1 || 1730 if (rq->curr->nr_cpus_allowed == 1 ||
2302 !cpudl_find(&rq->rd->cpudl, rq->c 1731 !cpudl_find(&rq->rd->cpudl, rq->curr, NULL))
2303 return; 1732 return;
2304 1733
2305 /* 1734 /*
2306 * p is migratable, so let's not sche 1735 * p is migratable, so let's not schedule it and
2307 * see if it is pushed or pulled some 1736 * see if it is pushed or pulled somewhere else.
2308 */ 1737 */
2309 if (p->nr_cpus_allowed != 1 && 1738 if (p->nr_cpus_allowed != 1 &&
2310 cpudl_find(&rq->rd->cpudl, p, NUL 1739 cpudl_find(&rq->rd->cpudl, p, NULL))
2311 return; 1740 return;
2312 1741
2313 resched_curr(rq); 1742 resched_curr(rq);
2314 } 1743 }
2315 1744
2316 static int balance_dl(struct rq *rq, struct t 1745 static int balance_dl(struct rq *rq, struct task_struct *p, struct rq_flags *rf)
2317 { 1746 {
2318 if (!on_dl_rq(&p->dl) && need_pull_dl 1747 if (!on_dl_rq(&p->dl) && need_pull_dl_task(rq, p)) {
2319 /* 1748 /*
2320 * This is OK, because curren 1749 * This is OK, because current is on_cpu, which avoids it being
2321 * picked for load-balance an 1750 * picked for load-balance and preemption/IRQs are still
2322 * disabled avoiding further 1751 * disabled avoiding further scheduler activity on it and we've
2323 * not yet started the pickin 1752 * not yet started the picking loop.
2324 */ 1753 */
2325 rq_unpin_lock(rq, rf); 1754 rq_unpin_lock(rq, rf);
2326 pull_dl_task(rq); 1755 pull_dl_task(rq);
2327 rq_repin_lock(rq, rf); 1756 rq_repin_lock(rq, rf);
2328 } 1757 }
2329 1758
2330 return sched_stop_runnable(rq) || sch 1759 return sched_stop_runnable(rq) || sched_dl_runnable(rq);
2331 } 1760 }
2332 #endif /* CONFIG_SMP */ 1761 #endif /* CONFIG_SMP */
2333 1762
2334 /* 1763 /*
2335 * Only called when both the current and waki 1764 * Only called when both the current and waking task are -deadline
2336 * tasks. 1765 * tasks.
2337 */ 1766 */
2338 static void wakeup_preempt_dl(struct rq *rq, !! 1767 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
2339 int flags) 1768 int flags)
2340 { 1769 {
2341 if (dl_entity_preempt(&p->dl, &rq->cu 1770 if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
2342 resched_curr(rq); 1771 resched_curr(rq);
2343 return; 1772 return;
2344 } 1773 }
2345 1774
2346 #ifdef CONFIG_SMP 1775 #ifdef CONFIG_SMP
2347 /* 1776 /*
2348 * In the unlikely case current and p 1777 * In the unlikely case current and p have the same deadline
2349 * let us try to decide what's the be 1778 * let us try to decide what's the best thing to do...
2350 */ 1779 */
2351 if ((p->dl.deadline == rq->curr->dl.d 1780 if ((p->dl.deadline == rq->curr->dl.deadline) &&
2352 !test_tsk_need_resched(rq->curr)) 1781 !test_tsk_need_resched(rq->curr))
2353 check_preempt_equal_dl(rq, p) 1782 check_preempt_equal_dl(rq, p);
2354 #endif /* CONFIG_SMP */ 1783 #endif /* CONFIG_SMP */
2355 } 1784 }
2356 1785
2357 #ifdef CONFIG_SCHED_HRTICK 1786 #ifdef CONFIG_SCHED_HRTICK
2358 static void start_hrtick_dl(struct rq *rq, st !! 1787 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
2359 { 1788 {
2360 hrtick_start(rq, dl_se->runtime); !! 1789 hrtick_start(rq, p->dl.runtime);
2361 } 1790 }
2362 #else /* !CONFIG_SCHED_HRTICK */ 1791 #else /* !CONFIG_SCHED_HRTICK */
2363 static void start_hrtick_dl(struct rq *rq, st !! 1792 static void start_hrtick_dl(struct rq *rq, struct task_struct *p)
2364 { 1793 {
2365 } 1794 }
2366 #endif 1795 #endif
2367 1796
2368 static void set_next_task_dl(struct rq *rq, s 1797 static void set_next_task_dl(struct rq *rq, struct task_struct *p, bool first)
2369 { 1798 {
2370 struct sched_dl_entity *dl_se = &p->d <<
2371 struct dl_rq *dl_rq = &rq->dl; <<
2372 <<
2373 p->se.exec_start = rq_clock_task(rq); 1799 p->se.exec_start = rq_clock_task(rq);
2374 if (on_dl_rq(&p->dl)) <<
2375 update_stats_wait_end_dl(dl_r <<
2376 1800
2377 /* You can't push away the running ta 1801 /* You can't push away the running task */
2378 dequeue_pushable_dl_task(rq, p); 1802 dequeue_pushable_dl_task(rq, p);
2379 1803
2380 if (!first) 1804 if (!first)
2381 return; 1805 return;
2382 1806
>> 1807 if (hrtick_enabled(rq))
>> 1808 start_hrtick_dl(rq, p);
>> 1809
2383 if (rq->curr->sched_class != &dl_sche 1810 if (rq->curr->sched_class != &dl_sched_class)
2384 update_dl_rq_load_avg(rq_cloc 1811 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 0);
2385 1812
2386 deadline_queue_push_tasks(rq); 1813 deadline_queue_push_tasks(rq);
2387 <<
2388 if (hrtick_enabled_dl(rq)) <<
2389 start_hrtick_dl(rq, &p->dl); <<
2390 } 1814 }
2391 1815
2392 static struct sched_dl_entity *pick_next_dl_e !! 1816 static struct sched_dl_entity *pick_next_dl_entity(struct rq *rq,
>> 1817 struct dl_rq *dl_rq)
2393 { 1818 {
2394 struct rb_node *left = rb_first_cache 1819 struct rb_node *left = rb_first_cached(&dl_rq->root);
2395 1820
2396 if (!left) 1821 if (!left)
2397 return NULL; 1822 return NULL;
2398 1823
2399 return __node_2_dle(left); !! 1824 return rb_entry(left, struct sched_dl_entity, rb_node);
2400 } 1825 }
2401 1826
2402 /* !! 1827 static struct task_struct *pick_next_task_dl(struct rq *rq)
2403 * __pick_next_task_dl - Helper to pick the n <<
2404 * @rq: The runqueue to pick the next task fr <<
2405 */ <<
2406 static struct task_struct *__pick_task_dl(str <<
2407 { 1828 {
2408 struct sched_dl_entity *dl_se; 1829 struct sched_dl_entity *dl_se;
2409 struct dl_rq *dl_rq = &rq->dl; 1830 struct dl_rq *dl_rq = &rq->dl;
2410 struct task_struct *p; 1831 struct task_struct *p;
2411 1832
2412 again: <<
2413 if (!sched_dl_runnable(rq)) 1833 if (!sched_dl_runnable(rq))
2414 return NULL; 1834 return NULL;
2415 1835
2416 dl_se = pick_next_dl_entity(dl_rq); !! 1836 dl_se = pick_next_dl_entity(rq, dl_rq);
2417 WARN_ON_ONCE(!dl_se); !! 1837 BUG_ON(!dl_se);
2418 !! 1838 p = dl_task_of(dl_se);
2419 if (dl_server(dl_se)) { !! 1839 set_next_task_dl(rq, p, true);
2420 p = dl_se->server_pick_task(d <<
2421 if (!p) { <<
2422 dl_se->dl_yielded = 1 <<
2423 update_curr_dl_se(rq, <<
2424 goto again; <<
2425 } <<
2426 rq->dl_server = dl_se; <<
2427 } else { <<
2428 p = dl_task_of(dl_se); <<
2429 } <<
2430 <<
2431 return p; 1840 return p;
2432 } 1841 }
2433 1842
2434 static struct task_struct *pick_task_dl(struc !! 1843 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
2435 { 1844 {
2436 return __pick_task_dl(rq); <<
2437 } <<
2438 <<
2439 static void put_prev_task_dl(struct rq *rq, s <<
2440 { <<
2441 struct sched_dl_entity *dl_se = &p->d <<
2442 struct dl_rq *dl_rq = &rq->dl; <<
2443 <<
2444 if (on_dl_rq(&p->dl)) <<
2445 update_stats_wait_start_dl(dl <<
2446 <<
2447 update_curr_dl(rq); 1845 update_curr_dl(rq);
2448 1846
2449 update_dl_rq_load_avg(rq_clock_pelt(r 1847 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2450 if (on_dl_rq(&p->dl) && p->nr_cpus_al 1848 if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
2451 enqueue_pushable_dl_task(rq, 1849 enqueue_pushable_dl_task(rq, p);
2452 } 1850 }
2453 1851
2454 /* 1852 /*
2455 * scheduler tick hitting a task of our sched 1853 * scheduler tick hitting a task of our scheduling class.
2456 * 1854 *
2457 * NOTE: This function can be called remotely 1855 * NOTE: This function can be called remotely by the tick offload that
2458 * goes along full dynticks. Therefore no loc 1856 * goes along full dynticks. Therefore no local assumption can be made
2459 * and everything must be accessed through th 1857 * and everything must be accessed through the @rq and @curr passed in
2460 * parameters. 1858 * parameters.
2461 */ 1859 */
2462 static void task_tick_dl(struct rq *rq, struc 1860 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
2463 { 1861 {
2464 update_curr_dl(rq); 1862 update_curr_dl(rq);
2465 1863
2466 update_dl_rq_load_avg(rq_clock_pelt(r 1864 update_dl_rq_load_avg(rq_clock_pelt(rq), rq, 1);
2467 /* 1865 /*
2468 * Even when we have runtime, update_ 1866 * Even when we have runtime, update_curr_dl() might have resulted in us
2469 * not being the leftmost task anymor 1867 * not being the leftmost task anymore. In that case NEED_RESCHED will
2470 * be set and schedule() will start a 1868 * be set and schedule() will start a new hrtick for the next task.
2471 */ 1869 */
2472 if (hrtick_enabled_dl(rq) && queued & !! 1870 if (hrtick_enabled(rq) && queued && p->dl.runtime > 0 &&
2473 is_leftmost(&p->dl, &rq->dl)) !! 1871 is_leftmost(p, &rq->dl))
2474 start_hrtick_dl(rq, &p->dl); !! 1872 start_hrtick_dl(rq, p);
2475 } 1873 }
2476 1874
2477 static void task_fork_dl(struct task_struct * 1875 static void task_fork_dl(struct task_struct *p)
2478 { 1876 {
2479 /* 1877 /*
2480 * SCHED_DEADLINE tasks cannot fork a 1878 * SCHED_DEADLINE tasks cannot fork and this is achieved through
2481 * sched_fork() 1879 * sched_fork()
2482 */ 1880 */
2483 } 1881 }
2484 1882
2485 #ifdef CONFIG_SMP 1883 #ifdef CONFIG_SMP
2486 1884
2487 /* Only try algorithms three times */ 1885 /* Only try algorithms three times */
2488 #define DL_MAX_TRIES 3 1886 #define DL_MAX_TRIES 3
2489 1887
2490 static int pick_dl_task(struct rq *rq, struct 1888 static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
2491 { 1889 {
2492 if (!task_on_cpu(rq, p) && !! 1890 if (!task_running(rq, p) &&
2493 cpumask_test_cpu(cpu, &p->cpus_ma !! 1891 cpumask_test_cpu(cpu, p->cpus_ptr))
2494 return 1; 1892 return 1;
2495 return 0; 1893 return 0;
2496 } 1894 }
2497 1895
2498 /* 1896 /*
2499 * Return the earliest pushable rq's task, wh 1897 * Return the earliest pushable rq's task, which is suitable to be executed
2500 * on the CPU, NULL otherwise: 1898 * on the CPU, NULL otherwise:
2501 */ 1899 */
2502 static struct task_struct *pick_earliest_push 1900 static struct task_struct *pick_earliest_pushable_dl_task(struct rq *rq, int cpu)
2503 { 1901 {
>> 1902 struct rb_node *next_node = rq->dl.pushable_dl_tasks_root.rb_leftmost;
2504 struct task_struct *p = NULL; 1903 struct task_struct *p = NULL;
2505 struct rb_node *next_node; <<
2506 1904
2507 if (!has_pushable_dl_tasks(rq)) 1905 if (!has_pushable_dl_tasks(rq))
2508 return NULL; 1906 return NULL;
2509 1907
2510 next_node = rb_first_cached(&rq->dl.p <<
2511 <<
2512 next_node: 1908 next_node:
2513 if (next_node) { 1909 if (next_node) {
2514 p = __node_2_pdl(next_node); !! 1910 p = rb_entry(next_node, struct task_struct, pushable_dl_tasks);
2515 1911
2516 if (pick_dl_task(rq, p, cpu)) 1912 if (pick_dl_task(rq, p, cpu))
2517 return p; 1913 return p;
2518 1914
2519 next_node = rb_next(next_node 1915 next_node = rb_next(next_node);
2520 goto next_node; 1916 goto next_node;
2521 } 1917 }
2522 1918
2523 return NULL; 1919 return NULL;
2524 } 1920 }
2525 1921
2526 static DEFINE_PER_CPU(cpumask_var_t, local_cp 1922 static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
2527 1923
2528 static int find_later_rq(struct task_struct * 1924 static int find_later_rq(struct task_struct *task)
2529 { 1925 {
2530 struct sched_domain *sd; 1926 struct sched_domain *sd;
2531 struct cpumask *later_mask = this_cpu 1927 struct cpumask *later_mask = this_cpu_cpumask_var_ptr(local_cpu_mask_dl);
2532 int this_cpu = smp_processor_id(); 1928 int this_cpu = smp_processor_id();
2533 int cpu = task_cpu(task); 1929 int cpu = task_cpu(task);
2534 1930
2535 /* Make sure the mask is initialized 1931 /* Make sure the mask is initialized first */
2536 if (unlikely(!later_mask)) 1932 if (unlikely(!later_mask))
2537 return -1; 1933 return -1;
2538 1934
2539 if (task->nr_cpus_allowed == 1) 1935 if (task->nr_cpus_allowed == 1)
2540 return -1; 1936 return -1;
2541 1937
2542 /* 1938 /*
2543 * We have to consider system topolog 1939 * We have to consider system topology and task affinity
2544 * first, then we can look for a suit 1940 * first, then we can look for a suitable CPU.
2545 */ 1941 */
2546 if (!cpudl_find(&task_rq(task)->rd->c 1942 if (!cpudl_find(&task_rq(task)->rd->cpudl, task, later_mask))
2547 return -1; 1943 return -1;
2548 1944
2549 /* 1945 /*
2550 * If we are here, some targets have 1946 * If we are here, some targets have been found, including
2551 * the most suitable which is, among 1947 * the most suitable which is, among the runqueues where the
2552 * current tasks have later deadlines 1948 * current tasks have later deadlines than the task's one, the
2553 * rq with the latest possible one. 1949 * rq with the latest possible one.
2554 * 1950 *
2555 * Now we check how well this matches 1951 * Now we check how well this matches with task's
2556 * affinity and system topology. 1952 * affinity and system topology.
2557 * 1953 *
2558 * The last CPU where the task run is 1954 * The last CPU where the task run is our first
2559 * guess, since it is most likely cac 1955 * guess, since it is most likely cache-hot there.
2560 */ 1956 */
2561 if (cpumask_test_cpu(cpu, later_mask) 1957 if (cpumask_test_cpu(cpu, later_mask))
2562 return cpu; 1958 return cpu;
2563 /* 1959 /*
2564 * Check if this_cpu is to be skipped 1960 * Check if this_cpu is to be skipped (i.e., it is
2565 * not in the mask) or not. 1961 * not in the mask) or not.
2566 */ 1962 */
2567 if (!cpumask_test_cpu(this_cpu, later 1963 if (!cpumask_test_cpu(this_cpu, later_mask))
2568 this_cpu = -1; 1964 this_cpu = -1;
2569 1965
2570 rcu_read_lock(); 1966 rcu_read_lock();
2571 for_each_domain(cpu, sd) { 1967 for_each_domain(cpu, sd) {
2572 if (sd->flags & SD_WAKE_AFFIN 1968 if (sd->flags & SD_WAKE_AFFINE) {
2573 int best_cpu; 1969 int best_cpu;
2574 1970
2575 /* 1971 /*
2576 * If possible, preem 1972 * If possible, preempting this_cpu is
2577 * cheaper than migra 1973 * cheaper than migrating.
2578 */ 1974 */
2579 if (this_cpu != -1 && 1975 if (this_cpu != -1 &&
2580 cpumask_test_cpu( 1976 cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
2581 rcu_read_unlo 1977 rcu_read_unlock();
2582 return this_c 1978 return this_cpu;
2583 } 1979 }
2584 1980
2585 best_cpu = cpumask_an !! 1981 best_cpu = cpumask_first_and(later_mask,
2586 !! 1982 sched_domain_span(sd));
2587 /* 1983 /*
2588 * Last chance: if a 1984 * Last chance: if a CPU being in both later_mask
2589 * and current sd spa 1985 * and current sd span is valid, that becomes our
2590 * choice. Of course, 1986 * choice. Of course, the latest possible CPU is
2591 * already under cons 1987 * already under consideration through later_mask.
2592 */ 1988 */
2593 if (best_cpu < nr_cpu 1989 if (best_cpu < nr_cpu_ids) {
2594 rcu_read_unlo 1990 rcu_read_unlock();
2595 return best_c 1991 return best_cpu;
2596 } 1992 }
2597 } 1993 }
2598 } 1994 }
2599 rcu_read_unlock(); 1995 rcu_read_unlock();
2600 1996
2601 /* 1997 /*
2602 * At this point, all our guesses fai 1998 * At this point, all our guesses failed, we just return
2603 * 'something', and let the caller so 1999 * 'something', and let the caller sort the things out.
2604 */ 2000 */
2605 if (this_cpu != -1) 2001 if (this_cpu != -1)
2606 return this_cpu; 2002 return this_cpu;
2607 2003
2608 cpu = cpumask_any_distribute(later_ma !! 2004 cpu = cpumask_any(later_mask);
2609 if (cpu < nr_cpu_ids) 2005 if (cpu < nr_cpu_ids)
2610 return cpu; 2006 return cpu;
2611 2007
2612 return -1; 2008 return -1;
2613 } 2009 }
2614 2010
2615 /* Locks the rq it finds */ 2011 /* Locks the rq it finds */
2616 static struct rq *find_lock_later_rq(struct t 2012 static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
2617 { 2013 {
2618 struct rq *later_rq = NULL; 2014 struct rq *later_rq = NULL;
2619 int tries; 2015 int tries;
2620 int cpu; 2016 int cpu;
2621 2017
2622 for (tries = 0; tries < DL_MAX_TRIES; 2018 for (tries = 0; tries < DL_MAX_TRIES; tries++) {
2623 cpu = find_later_rq(task); 2019 cpu = find_later_rq(task);
2624 2020
2625 if ((cpu == -1) || (cpu == rq 2021 if ((cpu == -1) || (cpu == rq->cpu))
2626 break; 2022 break;
2627 2023
2628 later_rq = cpu_rq(cpu); 2024 later_rq = cpu_rq(cpu);
2629 2025
2630 if (!dl_task_is_earliest_dead !! 2026 if (later_rq->dl.dl_nr_running &&
>> 2027 !dl_time_before(task->dl.deadline,
>> 2028 later_rq->dl.earliest_dl.curr)) {
2631 /* 2029 /*
2632 * Target rq has task 2030 * Target rq has tasks of equal or earlier deadline,
2633 * retrying does not 2031 * retrying does not release any lock and is unlikely
2634 * to yield a differe 2032 * to yield a different result.
2635 */ 2033 */
2636 later_rq = NULL; 2034 later_rq = NULL;
2637 break; 2035 break;
2638 } 2036 }
2639 2037
2640 /* Retry if something changed 2038 /* Retry if something changed. */
2641 if (double_lock_balance(rq, l 2039 if (double_lock_balance(rq, later_rq)) {
2642 if (unlikely(task_rq( 2040 if (unlikely(task_rq(task) != rq ||
2643 !cpumask !! 2041 !cpumask_test_cpu(later_rq->cpu, task->cpus_ptr) ||
2644 task_on_ !! 2042 task_running(rq, task) ||
2645 !dl_task 2043 !dl_task(task) ||
2646 is_migra <<
2647 !task_on 2044 !task_on_rq_queued(task))) {
2648 double_unlock 2045 double_unlock_balance(rq, later_rq);
2649 later_rq = NU 2046 later_rq = NULL;
2650 break; 2047 break;
2651 } 2048 }
2652 } 2049 }
2653 2050
2654 /* 2051 /*
2655 * If the rq we found has no 2052 * If the rq we found has no -deadline task, or
2656 * its earliest one has a lat 2053 * its earliest one has a later deadline than our
2657 * task, the rq is a good one 2054 * task, the rq is a good one.
2658 */ 2055 */
2659 if (dl_task_is_earliest_deadl !! 2056 if (!later_rq->dl.dl_nr_running ||
>> 2057 dl_time_before(task->dl.deadline,
>> 2058 later_rq->dl.earliest_dl.curr))
2660 break; 2059 break;
2661 2060
2662 /* Otherwise we try again. */ 2061 /* Otherwise we try again. */
2663 double_unlock_balance(rq, lat 2062 double_unlock_balance(rq, later_rq);
2664 later_rq = NULL; 2063 later_rq = NULL;
2665 } 2064 }
2666 2065
2667 return later_rq; 2066 return later_rq;
2668 } 2067 }
2669 2068
2670 static struct task_struct *pick_next_pushable 2069 static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
2671 { 2070 {
2672 struct task_struct *p; 2071 struct task_struct *p;
2673 2072
2674 if (!has_pushable_dl_tasks(rq)) 2073 if (!has_pushable_dl_tasks(rq))
2675 return NULL; 2074 return NULL;
2676 2075
2677 p = __node_2_pdl(rb_first_cached(&rq- !! 2076 p = rb_entry(rq->dl.pushable_dl_tasks_root.rb_leftmost,
>> 2077 struct task_struct, pushable_dl_tasks);
2678 2078
2679 WARN_ON_ONCE(rq->cpu != task_cpu(p)); !! 2079 BUG_ON(rq->cpu != task_cpu(p));
2680 WARN_ON_ONCE(task_current(rq, p)); !! 2080 BUG_ON(task_current(rq, p));
2681 WARN_ON_ONCE(p->nr_cpus_allowed <= 1) !! 2081 BUG_ON(p->nr_cpus_allowed <= 1);
2682 2082
2683 WARN_ON_ONCE(!task_on_rq_queued(p)); !! 2083 BUG_ON(!task_on_rq_queued(p));
2684 WARN_ON_ONCE(!dl_task(p)); !! 2084 BUG_ON(!dl_task(p));
2685 2085
2686 return p; 2086 return p;
2687 } 2087 }
2688 2088
2689 /* 2089 /*
2690 * See if the non running -deadline tasks on 2090 * See if the non running -deadline tasks on this rq
2691 * can be sent to some other CPU where they c 2091 * can be sent to some other CPU where they can preempt
2692 * and start executing. 2092 * and start executing.
2693 */ 2093 */
2694 static int push_dl_task(struct rq *rq) 2094 static int push_dl_task(struct rq *rq)
2695 { 2095 {
2696 struct task_struct *next_task; 2096 struct task_struct *next_task;
2697 struct rq *later_rq; 2097 struct rq *later_rq;
2698 int ret = 0; 2098 int ret = 0;
2699 2099
>> 2100 if (!rq->dl.overloaded)
>> 2101 return 0;
>> 2102
2700 next_task = pick_next_pushable_dl_tas 2103 next_task = pick_next_pushable_dl_task(rq);
2701 if (!next_task) 2104 if (!next_task)
2702 return 0; 2105 return 0;
2703 2106
2704 retry: 2107 retry:
>> 2108 if (WARN_ON(next_task == rq->curr))
>> 2109 return 0;
>> 2110
2705 /* 2111 /*
2706 * If next_task preempts rq->curr, an 2112 * If next_task preempts rq->curr, and rq->curr
2707 * can move away, it makes sense to j 2113 * can move away, it makes sense to just reschedule
2708 * without going further in pushing n 2114 * without going further in pushing next_task.
2709 */ 2115 */
2710 if (dl_task(rq->curr) && 2116 if (dl_task(rq->curr) &&
2711 dl_time_before(next_task->dl.dead 2117 dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
2712 rq->curr->nr_cpus_allowed > 1) { 2118 rq->curr->nr_cpus_allowed > 1) {
2713 resched_curr(rq); 2119 resched_curr(rq);
2714 return 0; 2120 return 0;
2715 } 2121 }
2716 2122
2717 if (is_migration_disabled(next_task)) <<
2718 return 0; <<
2719 <<
2720 if (WARN_ON(next_task == rq->curr)) <<
2721 return 0; <<
2722 <<
2723 /* We might release rq lock */ 2123 /* We might release rq lock */
2724 get_task_struct(next_task); 2124 get_task_struct(next_task);
2725 2125
2726 /* Will lock the rq it'll find */ 2126 /* Will lock the rq it'll find */
2727 later_rq = find_lock_later_rq(next_ta 2127 later_rq = find_lock_later_rq(next_task, rq);
2728 if (!later_rq) { 2128 if (!later_rq) {
2729 struct task_struct *task; 2129 struct task_struct *task;
2730 2130
2731 /* 2131 /*
2732 * We must check all this aga 2132 * We must check all this again, since
2733 * find_lock_later_rq release 2133 * find_lock_later_rq releases rq->lock and it is
2734 * then possible that next_ta 2134 * then possible that next_task has migrated.
2735 */ 2135 */
2736 task = pick_next_pushable_dl_ 2136 task = pick_next_pushable_dl_task(rq);
2737 if (task == next_task) { 2137 if (task == next_task) {
2738 /* 2138 /*
2739 * The task is still 2139 * The task is still there. We don't try
2740 * again, some other 2140 * again, some other CPU will pull it when ready.
2741 */ 2141 */
2742 goto out; 2142 goto out;
2743 } 2143 }
2744 2144
2745 if (!task) 2145 if (!task)
2746 /* No more tasks */ 2146 /* No more tasks */
2747 goto out; 2147 goto out;
2748 2148
2749 put_task_struct(next_task); 2149 put_task_struct(next_task);
2750 next_task = task; 2150 next_task = task;
2751 goto retry; 2151 goto retry;
2752 } 2152 }
2753 2153
2754 deactivate_task(rq, next_task, 0); 2154 deactivate_task(rq, next_task, 0);
2755 set_task_cpu(next_task, later_rq->cpu 2155 set_task_cpu(next_task, later_rq->cpu);
2756 activate_task(later_rq, next_task, 0) !! 2156
>> 2157 /*
>> 2158 * Update the later_rq clock here, because the clock is used
>> 2159 * by the cpufreq_update_util() inside __add_running_bw().
>> 2160 */
>> 2161 update_rq_clock(later_rq);
>> 2162 activate_task(later_rq, next_task, ENQUEUE_NOCLOCK);
2757 ret = 1; 2163 ret = 1;
2758 2164
2759 resched_curr(later_rq); 2165 resched_curr(later_rq);
2760 2166
2761 double_unlock_balance(rq, later_rq); 2167 double_unlock_balance(rq, later_rq);
2762 2168
2763 out: 2169 out:
2764 put_task_struct(next_task); 2170 put_task_struct(next_task);
2765 2171
2766 return ret; 2172 return ret;
2767 } 2173 }
2768 2174
2769 static void push_dl_tasks(struct rq *rq) 2175 static void push_dl_tasks(struct rq *rq)
2770 { 2176 {
2771 /* push_dl_task() will return true if 2177 /* push_dl_task() will return true if it moved a -deadline task */
2772 while (push_dl_task(rq)) 2178 while (push_dl_task(rq))
2773 ; 2179 ;
2774 } 2180 }
2775 2181
2776 static void pull_dl_task(struct rq *this_rq) 2182 static void pull_dl_task(struct rq *this_rq)
2777 { 2183 {
2778 int this_cpu = this_rq->cpu, cpu; 2184 int this_cpu = this_rq->cpu, cpu;
2779 struct task_struct *p, *push_task; !! 2185 struct task_struct *p;
2780 bool resched = false; 2186 bool resched = false;
2781 struct rq *src_rq; 2187 struct rq *src_rq;
2782 u64 dmin = LONG_MAX; 2188 u64 dmin = LONG_MAX;
2783 2189
2784 if (likely(!dl_overloaded(this_rq))) 2190 if (likely(!dl_overloaded(this_rq)))
2785 return; 2191 return;
2786 2192
2787 /* 2193 /*
2788 * Match the barrier from dl_set_over 2194 * Match the barrier from dl_set_overloaded; this guarantees that if we
2789 * see overloaded we must also see th 2195 * see overloaded we must also see the dlo_mask bit.
2790 */ 2196 */
2791 smp_rmb(); 2197 smp_rmb();
2792 2198
2793 for_each_cpu(cpu, this_rq->rd->dlo_ma 2199 for_each_cpu(cpu, this_rq->rd->dlo_mask) {
2794 if (this_cpu == cpu) 2200 if (this_cpu == cpu)
2795 continue; 2201 continue;
2796 2202
2797 src_rq = cpu_rq(cpu); 2203 src_rq = cpu_rq(cpu);
2798 2204
2799 /* 2205 /*
2800 * It looks racy, and it is! !! 2206 * It looks racy, abd it is! However, as in sched_rt.c,
2801 * we are fine with this. 2207 * we are fine with this.
2802 */ 2208 */
2803 if (this_rq->dl.dl_nr_running 2209 if (this_rq->dl.dl_nr_running &&
2804 dl_time_before(this_rq->d 2210 dl_time_before(this_rq->dl.earliest_dl.curr,
2805 src_rq->dl 2211 src_rq->dl.earliest_dl.next))
2806 continue; 2212 continue;
2807 2213
2808 /* Might drop this_rq->lock * 2214 /* Might drop this_rq->lock */
2809 push_task = NULL; <<
2810 double_lock_balance(this_rq, 2215 double_lock_balance(this_rq, src_rq);
2811 2216
2812 /* 2217 /*
2813 * If there are no more pulla 2218 * If there are no more pullable tasks on the
2814 * rq, we're done with it. 2219 * rq, we're done with it.
2815 */ 2220 */
2816 if (src_rq->dl.dl_nr_running 2221 if (src_rq->dl.dl_nr_running <= 1)
2817 goto skip; 2222 goto skip;
2818 2223
2819 p = pick_earliest_pushable_dl 2224 p = pick_earliest_pushable_dl_task(src_rq, this_cpu);
2820 2225
2821 /* 2226 /*
2822 * We found a task to be pull 2227 * We found a task to be pulled if:
2823 * - it preempts our current 2228 * - it preempts our current (if there's one),
2824 * - it will preempt the las 2229 * - it will preempt the last one we pulled (if any).
2825 */ 2230 */
2826 if (p && dl_time_before(p->dl 2231 if (p && dl_time_before(p->dl.deadline, dmin) &&
2827 dl_task_is_earliest_deadl !! 2232 (!this_rq->dl.dl_nr_running ||
>> 2233 dl_time_before(p->dl.deadline,
>> 2234 this_rq->dl.earliest_dl.curr))) {
2828 WARN_ON(p == src_rq-> 2235 WARN_ON(p == src_rq->curr);
2829 WARN_ON(!task_on_rq_q 2236 WARN_ON(!task_on_rq_queued(p));
2830 2237
2831 /* 2238 /*
2832 * Then we pull iff p 2239 * Then we pull iff p has actually an earlier
2833 * deadline than the 2240 * deadline than the current task of its runqueue.
2834 */ 2241 */
2835 if (dl_time_before(p- 2242 if (dl_time_before(p->dl.deadline,
2836 sr 2243 src_rq->curr->dl.deadline))
2837 goto skip; 2244 goto skip;
2838 2245
2839 if (is_migration_disa !! 2246 resched = true;
2840 push_task = g !! 2247
2841 } else { !! 2248 deactivate_task(src_rq, p, 0);
2842 deactivate_ta !! 2249 set_task_cpu(p, this_cpu);
2843 set_task_cpu( !! 2250 activate_task(this_rq, p, 0);
2844 activate_task !! 2251 dmin = p->dl.deadline;
2845 dmin = p->dl. <<
2846 resched = tru <<
2847 } <<
2848 2252
2849 /* Is there any other 2253 /* Is there any other task even earlier? */
2850 } 2254 }
2851 skip: 2255 skip:
2852 double_unlock_balance(this_rq 2256 double_unlock_balance(this_rq, src_rq);
2853 <<
2854 if (push_task) { <<
2855 preempt_disable(); <<
2856 raw_spin_rq_unlock(th <<
2857 stop_one_cpu_nowait(s <<
2858 p <<
2859 preempt_enable(); <<
2860 raw_spin_rq_lock(this <<
2861 } <<
2862 } 2257 }
2863 2258
2864 if (resched) 2259 if (resched)
2865 resched_curr(this_rq); 2260 resched_curr(this_rq);
2866 } 2261 }
2867 2262
2868 /* 2263 /*
2869 * Since the task is not running and a resche 2264 * Since the task is not running and a reschedule is not going to happen
2870 * anytime soon on its runqueue, we try pushi 2265 * anytime soon on its runqueue, we try pushing it away now.
2871 */ 2266 */
2872 static void task_woken_dl(struct rq *rq, stru 2267 static void task_woken_dl(struct rq *rq, struct task_struct *p)
2873 { 2268 {
2874 if (!task_on_cpu(rq, p) && !! 2269 if (!task_running(rq, p) &&
2875 !test_tsk_need_resched(rq->curr) 2270 !test_tsk_need_resched(rq->curr) &&
2876 p->nr_cpus_allowed > 1 && 2271 p->nr_cpus_allowed > 1 &&
2877 dl_task(rq->curr) && 2272 dl_task(rq->curr) &&
2878 (rq->curr->nr_cpus_allowed < 2 || 2273 (rq->curr->nr_cpus_allowed < 2 ||
2879 !dl_entity_preempt(&p->dl, &rq-> 2274 !dl_entity_preempt(&p->dl, &rq->curr->dl))) {
2880 push_dl_tasks(rq); 2275 push_dl_tasks(rq);
2881 } 2276 }
2882 } 2277 }
2883 2278
2884 static void set_cpus_allowed_dl(struct task_s 2279 static void set_cpus_allowed_dl(struct task_struct *p,
2885 struct affini !! 2280 const struct cpumask *new_mask)
2886 { 2281 {
2887 struct root_domain *src_rd; 2282 struct root_domain *src_rd;
2888 struct rq *rq; 2283 struct rq *rq;
2889 2284
2890 WARN_ON_ONCE(!dl_task(p)); !! 2285 BUG_ON(!dl_task(p));
2891 2286
2892 rq = task_rq(p); 2287 rq = task_rq(p);
2893 src_rd = rq->rd; 2288 src_rd = rq->rd;
2894 /* 2289 /*
2895 * Migrating a SCHED_DEADLINE task be 2290 * Migrating a SCHED_DEADLINE task between exclusive
2896 * cpusets (different root_domains) e 2291 * cpusets (different root_domains) entails a bandwidth
2897 * update. We already made space for 2292 * update. We already made space for us in the destination
2898 * domain (see cpuset_can_attach()). 2293 * domain (see cpuset_can_attach()).
2899 */ 2294 */
2900 if (!cpumask_intersects(src_rd->span, !! 2295 if (!cpumask_intersects(src_rd->span, new_mask)) {
2901 struct dl_bw *src_dl_b; 2296 struct dl_bw *src_dl_b;
2902 2297
2903 src_dl_b = dl_bw_of(cpu_of(rq 2298 src_dl_b = dl_bw_of(cpu_of(rq));
2904 /* 2299 /*
2905 * We now free resources of t 2300 * We now free resources of the root_domain we are migrating
2906 * off. In the worst case, sc 2301 * off. In the worst case, sched_setattr() may temporary fail
2907 * until we complete the upda 2302 * until we complete the update.
2908 */ 2303 */
2909 raw_spin_lock(&src_dl_b->lock 2304 raw_spin_lock(&src_dl_b->lock);
2910 __dl_sub(src_dl_b, p->dl.dl_b 2305 __dl_sub(src_dl_b, p->dl.dl_bw, dl_bw_cpus(task_cpu(p)));
2911 raw_spin_unlock(&src_dl_b->lo 2306 raw_spin_unlock(&src_dl_b->lock);
2912 } 2307 }
2913 2308
2914 set_cpus_allowed_common(p, ctx); !! 2309 set_cpus_allowed_common(p, new_mask);
2915 } 2310 }
2916 2311
2917 /* Assumes rq->lock is held */ 2312 /* Assumes rq->lock is held */
2918 static void rq_online_dl(struct rq *rq) 2313 static void rq_online_dl(struct rq *rq)
2919 { 2314 {
2920 if (rq->dl.overloaded) 2315 if (rq->dl.overloaded)
2921 dl_set_overload(rq); 2316 dl_set_overload(rq);
2922 2317
2923 cpudl_set_freecpu(&rq->rd->cpudl, rq- 2318 cpudl_set_freecpu(&rq->rd->cpudl, rq->cpu);
2924 if (rq->dl.dl_nr_running > 0) 2319 if (rq->dl.dl_nr_running > 0)
2925 cpudl_set(&rq->rd->cpudl, rq- 2320 cpudl_set(&rq->rd->cpudl, rq->cpu, rq->dl.earliest_dl.curr);
2926 } 2321 }
2927 2322
2928 /* Assumes rq->lock is held */ 2323 /* Assumes rq->lock is held */
2929 static void rq_offline_dl(struct rq *rq) 2324 static void rq_offline_dl(struct rq *rq)
2930 { 2325 {
2931 if (rq->dl.overloaded) 2326 if (rq->dl.overloaded)
2932 dl_clear_overload(rq); 2327 dl_clear_overload(rq);
2933 2328
2934 cpudl_clear(&rq->rd->cpudl, rq->cpu); 2329 cpudl_clear(&rq->rd->cpudl, rq->cpu);
2935 cpudl_clear_freecpu(&rq->rd->cpudl, r 2330 cpudl_clear_freecpu(&rq->rd->cpudl, rq->cpu);
2936 } 2331 }
2937 2332
2938 void __init init_sched_dl_class(void) 2333 void __init init_sched_dl_class(void)
2939 { 2334 {
2940 unsigned int i; 2335 unsigned int i;
2941 2336
2942 for_each_possible_cpu(i) 2337 for_each_possible_cpu(i)
2943 zalloc_cpumask_var_node(&per_ 2338 zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
2944 GFP_K 2339 GFP_KERNEL, cpu_to_node(i));
2945 } 2340 }
2946 2341
2947 void dl_add_task_root_domain(struct task_stru 2342 void dl_add_task_root_domain(struct task_struct *p)
2948 { 2343 {
2949 struct rq_flags rf; 2344 struct rq_flags rf;
2950 struct rq *rq; 2345 struct rq *rq;
2951 struct dl_bw *dl_b; 2346 struct dl_bw *dl_b;
2952 2347
2953 raw_spin_lock_irqsave(&p->pi_lock, rf !! 2348 rq = task_rq_lock(p, &rf);
2954 if (!dl_task(p)) { !! 2349 if (!dl_task(p))
2955 raw_spin_unlock_irqrestore(&p !! 2350 goto unlock;
2956 return; <<
2957 } <<
2958 <<
2959 rq = __task_rq_lock(p, &rf); <<
2960 2351
2961 dl_b = &rq->rd->dl_bw; 2352 dl_b = &rq->rd->dl_bw;
2962 raw_spin_lock(&dl_b->lock); 2353 raw_spin_lock(&dl_b->lock);
2963 2354
2964 __dl_add(dl_b, p->dl.dl_bw, cpumask_w 2355 __dl_add(dl_b, p->dl.dl_bw, cpumask_weight(rq->rd->span));
2965 2356
2966 raw_spin_unlock(&dl_b->lock); 2357 raw_spin_unlock(&dl_b->lock);
2967 2358
>> 2359 unlock:
2968 task_rq_unlock(rq, p, &rf); 2360 task_rq_unlock(rq, p, &rf);
2969 } 2361 }
2970 2362
2971 void dl_clear_root_domain(struct root_domain 2363 void dl_clear_root_domain(struct root_domain *rd)
2972 { 2364 {
2973 unsigned long flags; 2365 unsigned long flags;
2974 2366
2975 raw_spin_lock_irqsave(&rd->dl_bw.lock 2367 raw_spin_lock_irqsave(&rd->dl_bw.lock, flags);
2976 rd->dl_bw.total_bw = 0; 2368 rd->dl_bw.total_bw = 0;
2977 raw_spin_unlock_irqrestore(&rd->dl_bw 2369 raw_spin_unlock_irqrestore(&rd->dl_bw.lock, flags);
2978 } 2370 }
2979 2371
2980 #endif /* CONFIG_SMP */ 2372 #endif /* CONFIG_SMP */
2981 2373
2982 static void switched_from_dl(struct rq *rq, s 2374 static void switched_from_dl(struct rq *rq, struct task_struct *p)
2983 { 2375 {
2984 /* 2376 /*
2985 * task_non_contending() can start th 2377 * task_non_contending() can start the "inactive timer" (if the 0-lag
2986 * time is in the future). If the tas 2378 * time is in the future). If the task switches back to dl before
2987 * the "inactive timer" fires, it can 2379 * the "inactive timer" fires, it can continue to consume its current
2988 * runtime using its current deadline 2380 * runtime using its current deadline. If it stays outside of
2989 * SCHED_DEADLINE until the 0-lag tim 2381 * SCHED_DEADLINE until the 0-lag time passes, inactive_task_timer()
2990 * will reset the task parameters. 2382 * will reset the task parameters.
2991 */ 2383 */
2992 if (task_on_rq_queued(p) && p->dl.dl_ 2384 if (task_on_rq_queued(p) && p->dl.dl_runtime)
2993 task_non_contending(&p->dl); !! 2385 task_non_contending(p);
2994 <<
2995 /* <<
2996 * In case a task is setscheduled out <<
2997 * keep track of that on its cpuset ( <<
2998 */ <<
2999 dec_dl_tasks_cs(p); <<
3000 2386
3001 if (!task_on_rq_queued(p)) { 2387 if (!task_on_rq_queued(p)) {
3002 /* 2388 /*
3003 * Inactive timer is armed. H 2389 * Inactive timer is armed. However, p is leaving DEADLINE and
3004 * might migrate away from th 2390 * might migrate away from this rq while continuing to run on
3005 * some other class. We need 2391 * some other class. We need to remove its contribution from
3006 * this rq running_bw now, or 2392 * this rq running_bw now, or sub_rq_bw (below) will complain.
3007 */ 2393 */
3008 if (p->dl.dl_non_contending) 2394 if (p->dl.dl_non_contending)
3009 sub_running_bw(&p->dl 2395 sub_running_bw(&p->dl, &rq->dl);
3010 sub_rq_bw(&p->dl, &rq->dl); 2396 sub_rq_bw(&p->dl, &rq->dl);
3011 } 2397 }
3012 2398
3013 /* 2399 /*
3014 * We cannot use inactive_task_timer( 2400 * We cannot use inactive_task_timer() to invoke sub_running_bw()
3015 * at the 0-lag time, because the tas 2401 * at the 0-lag time, because the task could have been migrated
3016 * while SCHED_OTHER in the meanwhile 2402 * while SCHED_OTHER in the meanwhile.
3017 */ 2403 */
3018 if (p->dl.dl_non_contending) 2404 if (p->dl.dl_non_contending)
3019 p->dl.dl_non_contending = 0; 2405 p->dl.dl_non_contending = 0;
3020 2406
3021 /* 2407 /*
3022 * Since this might be the only -dead 2408 * Since this might be the only -deadline task on the rq,
3023 * this is the right place to try to 2409 * this is the right place to try to pull some other one
3024 * from an overloaded CPU, if any. 2410 * from an overloaded CPU, if any.
3025 */ 2411 */
3026 if (!task_on_rq_queued(p) || rq->dl.d 2412 if (!task_on_rq_queued(p) || rq->dl.dl_nr_running)
3027 return; 2413 return;
3028 2414
3029 deadline_queue_pull_task(rq); 2415 deadline_queue_pull_task(rq);
3030 } 2416 }
3031 2417
3032 /* 2418 /*
3033 * When switching to -deadline, we may overlo 2419 * When switching to -deadline, we may overload the rq, then
3034 * we try to push someone off, if possible. 2420 * we try to push someone off, if possible.
3035 */ 2421 */
3036 static void switched_to_dl(struct rq *rq, str 2422 static void switched_to_dl(struct rq *rq, struct task_struct *p)
3037 { 2423 {
3038 if (hrtimer_try_to_cancel(&p->dl.inac 2424 if (hrtimer_try_to_cancel(&p->dl.inactive_timer) == 1)
3039 put_task_struct(p); 2425 put_task_struct(p);
3040 2426
3041 /* <<
3042 * In case a task is setscheduled to <<
3043 * track of that on its cpuset (for c <<
3044 */ <<
3045 inc_dl_tasks_cs(p); <<
3046 <<
3047 /* If p is not queued we will update 2427 /* If p is not queued we will update its parameters at next wakeup. */
3048 if (!task_on_rq_queued(p)) { 2428 if (!task_on_rq_queued(p)) {
3049 add_rq_bw(&p->dl, &rq->dl); 2429 add_rq_bw(&p->dl, &rq->dl);
3050 2430
3051 return; 2431 return;
3052 } 2432 }
3053 2433
3054 if (rq->curr != p) { 2434 if (rq->curr != p) {
3055 #ifdef CONFIG_SMP 2435 #ifdef CONFIG_SMP
3056 if (p->nr_cpus_allowed > 1 && 2436 if (p->nr_cpus_allowed > 1 && rq->dl.overloaded)
3057 deadline_queue_push_t 2437 deadline_queue_push_tasks(rq);
3058 #endif 2438 #endif
3059 if (dl_task(rq->curr)) 2439 if (dl_task(rq->curr))
3060 wakeup_preempt_dl(rq, !! 2440 check_preempt_curr_dl(rq, p, 0);
3061 else 2441 else
3062 resched_curr(rq); 2442 resched_curr(rq);
3063 } else { <<
3064 update_dl_rq_load_avg(rq_cloc <<
3065 } 2443 }
3066 } 2444 }
3067 2445
3068 /* 2446 /*
3069 * If the scheduling parameters of a -deadlin 2447 * If the scheduling parameters of a -deadline task changed,
3070 * a push or pull operation might be needed. 2448 * a push or pull operation might be needed.
3071 */ 2449 */
3072 static void prio_changed_dl(struct rq *rq, st 2450 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
3073 int oldprio) 2451 int oldprio)
3074 { 2452 {
3075 if (!task_on_rq_queued(p)) !! 2453 if (task_on_rq_queued(p) || rq->curr == p) {
3076 return; <<
3077 <<
3078 #ifdef CONFIG_SMP 2454 #ifdef CONFIG_SMP
3079 /* !! 2455 /*
3080 * This might be too much, but unfort !! 2456 * This might be too much, but unfortunately
3081 * we don't have the old deadline val !! 2457 * we don't have the old deadline value, and
3082 * we can't argue if the task is incr !! 2458 * we can't argue if the task is increasing
3083 * or lowering its prio, so... !! 2459 * or lowering its prio, so...
3084 */ !! 2460 */
3085 if (!rq->dl.overloaded) !! 2461 if (!rq->dl.overloaded)
3086 deadline_queue_pull_task(rq); !! 2462 deadline_queue_pull_task(rq);
3087 2463
3088 if (task_current(rq, p)) { <<
3089 /* 2464 /*
3090 * If we now have a earlier d 2465 * If we now have a earlier deadline task than p,
3091 * then reschedule, provided 2466 * then reschedule, provided p is still on this
3092 * runqueue. 2467 * runqueue.
3093 */ 2468 */
3094 if (dl_time_before(rq->dl.ear 2469 if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline))
3095 resched_curr(rq); 2470 resched_curr(rq);
3096 } else { !! 2471 #else
3097 /* 2472 /*
3098 * Current may not be deadlin !! 2473 * Again, we don't know if p has a earlier
3099 * have just replenished it ( !! 2474 * or later deadline, so let's blindly set a
3100 * !! 2475 * (maybe not needed) rescheduling point.
3101 * Otherwise, if p was given <<
3102 */ 2476 */
3103 if (!dl_task(rq->curr) || !! 2477 resched_curr(rq);
3104 dl_time_before(p->dl.dead !! 2478 #endif /* CONFIG_SMP */
3105 resched_curr(rq); <<
3106 } 2479 }
3107 #else <<
3108 /* <<
3109 * We don't know if p has a earlier o <<
3110 * set a (maybe not needed) reschedul <<
3111 */ <<
3112 resched_curr(rq); <<
3113 #endif <<
3114 } <<
3115 <<
3116 #ifdef CONFIG_SCHED_CORE <<
3117 static int task_is_throttled_dl(struct task_s <<
3118 { <<
3119 return p->dl.dl_throttled; <<
3120 } 2480 }
3121 #endif <<
3122 <<
3123 DEFINE_SCHED_CLASS(dl) = { <<
3124 2481
>> 2482 const struct sched_class dl_sched_class
>> 2483 __attribute__((section("__dl_sched_class"))) = {
3125 .enqueue_task = enqueue_tas 2484 .enqueue_task = enqueue_task_dl,
3126 .dequeue_task = dequeue_tas 2485 .dequeue_task = dequeue_task_dl,
3127 .yield_task = yield_task_ 2486 .yield_task = yield_task_dl,
3128 2487
3129 .wakeup_preempt = wakeup_pree !! 2488 .check_preempt_curr = check_preempt_curr_dl,
3130 2489
3131 .pick_task = pick_task_d !! 2490 .pick_next_task = pick_next_task_dl,
3132 .put_prev_task = put_prev_ta 2491 .put_prev_task = put_prev_task_dl,
3133 .set_next_task = set_next_ta 2492 .set_next_task = set_next_task_dl,
3134 2493
3135 #ifdef CONFIG_SMP 2494 #ifdef CONFIG_SMP
3136 .balance = balance_dl, 2495 .balance = balance_dl,
3137 .select_task_rq = select_task 2496 .select_task_rq = select_task_rq_dl,
3138 .migrate_task_rq = migrate_tas 2497 .migrate_task_rq = migrate_task_rq_dl,
3139 .set_cpus_allowed = set_cpus_al 2498 .set_cpus_allowed = set_cpus_allowed_dl,
3140 .rq_online = rq_online_d 2499 .rq_online = rq_online_dl,
3141 .rq_offline = rq_offline_ 2500 .rq_offline = rq_offline_dl,
3142 .task_woken = task_woken_ 2501 .task_woken = task_woken_dl,
3143 .find_lock_rq = find_lock_l <<
3144 #endif 2502 #endif
3145 2503
3146 .task_tick = task_tick_d 2504 .task_tick = task_tick_dl,
3147 .task_fork = task_fork_d 2505 .task_fork = task_fork_dl,
3148 2506
3149 .prio_changed = prio_change 2507 .prio_changed = prio_changed_dl,
3150 .switched_from = switched_fr 2508 .switched_from = switched_from_dl,
3151 .switched_to = switched_to 2509 .switched_to = switched_to_dl,
3152 2510
3153 .update_curr = update_curr 2511 .update_curr = update_curr_dl,
3154 #ifdef CONFIG_SCHED_CORE <<
3155 .task_is_throttled = task_is_thr <<
3156 #endif <<
3157 }; 2512 };
3158 2513
3159 /* Used for dl_bw check and update, used unde <<
3160 static u64 dl_generation; <<
3161 <<
3162 int sched_dl_global_validate(void) 2514 int sched_dl_global_validate(void)
3163 { 2515 {
3164 u64 runtime = global_rt_runtime(); 2516 u64 runtime = global_rt_runtime();
3165 u64 period = global_rt_period(); 2517 u64 period = global_rt_period();
3166 u64 new_bw = to_ratio(period, runtime 2518 u64 new_bw = to_ratio(period, runtime);
3167 u64 gen = ++dl_generation; <<
3168 struct dl_bw *dl_b; 2519 struct dl_bw *dl_b;
3169 int cpu, cpus, ret = 0; !! 2520 int cpu, ret = 0;
3170 unsigned long flags; 2521 unsigned long flags;
3171 2522
3172 /* 2523 /*
3173 * Here we want to check the bandwidt 2524 * Here we want to check the bandwidth not being set to some
3174 * value smaller than the currently a 2525 * value smaller than the currently allocated bandwidth in
3175 * any of the root_domains. 2526 * any of the root_domains.
>> 2527 *
>> 2528 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
>> 2529 * cycling on root_domains... Discussion on different/better
>> 2530 * solutions is welcome!
3176 */ 2531 */
3177 for_each_possible_cpu(cpu) { 2532 for_each_possible_cpu(cpu) {
3178 rcu_read_lock_sched(); 2533 rcu_read_lock_sched();
3179 <<
3180 if (dl_bw_visited(cpu, gen)) <<
3181 goto next; <<
3182 <<
3183 dl_b = dl_bw_of(cpu); 2534 dl_b = dl_bw_of(cpu);
3184 cpus = dl_bw_cpus(cpu); <<
3185 2535
3186 raw_spin_lock_irqsave(&dl_b-> 2536 raw_spin_lock_irqsave(&dl_b->lock, flags);
3187 if (new_bw * cpus < dl_b->tot !! 2537 if (new_bw < dl_b->total_bw)
3188 ret = -EBUSY; 2538 ret = -EBUSY;
3189 raw_spin_unlock_irqrestore(&d 2539 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3190 2540
3191 next: <<
3192 rcu_read_unlock_sched(); 2541 rcu_read_unlock_sched();
3193 2542
3194 if (ret) 2543 if (ret)
3195 break; 2544 break;
3196 } 2545 }
3197 2546
3198 return ret; 2547 return ret;
3199 } 2548 }
3200 2549
3201 static void init_dl_rq_bw_ratio(struct dl_rq 2550 static void init_dl_rq_bw_ratio(struct dl_rq *dl_rq)
3202 { 2551 {
3203 if (global_rt_runtime() == RUNTIME_IN 2552 if (global_rt_runtime() == RUNTIME_INF) {
3204 dl_rq->bw_ratio = 1 << RATIO_ 2553 dl_rq->bw_ratio = 1 << RATIO_SHIFT;
3205 dl_rq->max_bw = dl_rq->extra_ !! 2554 dl_rq->extra_bw = 1 << BW_SHIFT;
3206 } else { 2555 } else {
3207 dl_rq->bw_ratio = to_ratio(gl 2556 dl_rq->bw_ratio = to_ratio(global_rt_runtime(),
3208 global_rt_period()) 2557 global_rt_period()) >> (BW_SHIFT - RATIO_SHIFT);
3209 dl_rq->max_bw = dl_rq->extra_ !! 2558 dl_rq->extra_bw = to_ratio(global_rt_period(),
3210 to_ratio(global_rt_pe !! 2559 global_rt_runtime());
3211 } 2560 }
3212 } 2561 }
3213 2562
3214 void sched_dl_do_global(void) 2563 void sched_dl_do_global(void)
3215 { 2564 {
3216 u64 new_bw = -1; 2565 u64 new_bw = -1;
3217 u64 gen = ++dl_generation; <<
3218 struct dl_bw *dl_b; 2566 struct dl_bw *dl_b;
3219 int cpu; 2567 int cpu;
3220 unsigned long flags; 2568 unsigned long flags;
3221 2569
>> 2570 def_dl_bandwidth.dl_period = global_rt_period();
>> 2571 def_dl_bandwidth.dl_runtime = global_rt_runtime();
>> 2572
3222 if (global_rt_runtime() != RUNTIME_IN 2573 if (global_rt_runtime() != RUNTIME_INF)
3223 new_bw = to_ratio(global_rt_p 2574 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
3224 2575
>> 2576 /*
>> 2577 * FIXME: As above...
>> 2578 */
3225 for_each_possible_cpu(cpu) { 2579 for_each_possible_cpu(cpu) {
3226 rcu_read_lock_sched(); 2580 rcu_read_lock_sched();
3227 <<
3228 if (dl_bw_visited(cpu, gen)) <<
3229 rcu_read_unlock_sched <<
3230 continue; <<
3231 } <<
3232 <<
3233 dl_b = dl_bw_of(cpu); 2581 dl_b = dl_bw_of(cpu);
3234 2582
3235 raw_spin_lock_irqsave(&dl_b-> 2583 raw_spin_lock_irqsave(&dl_b->lock, flags);
3236 dl_b->bw = new_bw; 2584 dl_b->bw = new_bw;
3237 raw_spin_unlock_irqrestore(&d 2585 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3238 2586
3239 rcu_read_unlock_sched(); 2587 rcu_read_unlock_sched();
3240 init_dl_rq_bw_ratio(&cpu_rq(c 2588 init_dl_rq_bw_ratio(&cpu_rq(cpu)->dl);
3241 } 2589 }
3242 } 2590 }
3243 2591
3244 /* 2592 /*
3245 * We must be sure that accepting a new task 2593 * We must be sure that accepting a new task (or allowing changing the
3246 * parameters of an existing one) is consiste 2594 * parameters of an existing one) is consistent with the bandwidth
3247 * constraints. If yes, this function also ac 2595 * constraints. If yes, this function also accordingly updates the currently
3248 * allocated bandwidth to reflect the new sit 2596 * allocated bandwidth to reflect the new situation.
3249 * 2597 *
3250 * This function is called while holding p's 2598 * This function is called while holding p's rq->lock.
3251 */ 2599 */
3252 int sched_dl_overflow(struct task_struct *p, 2600 int sched_dl_overflow(struct task_struct *p, int policy,
3253 const struct sched_attr 2601 const struct sched_attr *attr)
3254 { 2602 {
3255 u64 period = attr->sched_period ?: at 2603 u64 period = attr->sched_period ?: attr->sched_deadline;
3256 u64 runtime = attr->sched_runtime; 2604 u64 runtime = attr->sched_runtime;
3257 u64 new_bw = dl_policy(policy) ? to_r 2605 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
3258 int cpus, err = -1, cpu = task_cpu(p) 2606 int cpus, err = -1, cpu = task_cpu(p);
3259 struct dl_bw *dl_b = dl_bw_of(cpu); 2607 struct dl_bw *dl_b = dl_bw_of(cpu);
3260 unsigned long cap; 2608 unsigned long cap;
3261 2609
3262 if (attr->sched_flags & SCHED_FLAG_SU 2610 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3263 return 0; 2611 return 0;
3264 2612
3265 /* !deadline task may carry old deadl 2613 /* !deadline task may carry old deadline bandwidth */
3266 if (new_bw == p->dl.dl_bw && task_has 2614 if (new_bw == p->dl.dl_bw && task_has_dl_policy(p))
3267 return 0; 2615 return 0;
3268 2616
3269 /* 2617 /*
3270 * Either if a task, enters, leave, o 2618 * Either if a task, enters, leave, or stays -deadline but changes
3271 * its parameters, we may need to upd 2619 * its parameters, we may need to update accordingly the total
3272 * allocated bandwidth of the contain 2620 * allocated bandwidth of the container.
3273 */ 2621 */
3274 raw_spin_lock(&dl_b->lock); 2622 raw_spin_lock(&dl_b->lock);
3275 cpus = dl_bw_cpus(cpu); 2623 cpus = dl_bw_cpus(cpu);
3276 cap = dl_bw_capacity(cpu); 2624 cap = dl_bw_capacity(cpu);
3277 2625
3278 if (dl_policy(policy) && !task_has_dl 2626 if (dl_policy(policy) && !task_has_dl_policy(p) &&
3279 !__dl_overflow(dl_b, cap, 0, new_ 2627 !__dl_overflow(dl_b, cap, 0, new_bw)) {
3280 if (hrtimer_active(&p->dl.ina 2628 if (hrtimer_active(&p->dl.inactive_timer))
3281 __dl_sub(dl_b, p->dl. 2629 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3282 __dl_add(dl_b, new_bw, cpus); 2630 __dl_add(dl_b, new_bw, cpus);
3283 err = 0; 2631 err = 0;
3284 } else if (dl_policy(policy) && task_ 2632 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
3285 !__dl_overflow(dl_b, cap, 2633 !__dl_overflow(dl_b, cap, p->dl.dl_bw, new_bw)) {
3286 /* 2634 /*
3287 * XXX this is slightly incor 2635 * XXX this is slightly incorrect: when the task
3288 * utilization decreases, we 2636 * utilization decreases, we should delay the total
3289 * utilization change until t 2637 * utilization change until the task's 0-lag point.
3290 * But this would require to 2638 * But this would require to set the task's "inactive
3291 * timer" when the task is no 2639 * timer" when the task is not inactive.
3292 */ 2640 */
3293 __dl_sub(dl_b, p->dl.dl_bw, c 2641 __dl_sub(dl_b, p->dl.dl_bw, cpus);
3294 __dl_add(dl_b, new_bw, cpus); 2642 __dl_add(dl_b, new_bw, cpus);
3295 dl_change_utilization(p, new_ 2643 dl_change_utilization(p, new_bw);
3296 err = 0; 2644 err = 0;
3297 } else if (!dl_policy(policy) && task 2645 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
3298 /* 2646 /*
3299 * Do not decrease the total 2647 * Do not decrease the total deadline utilization here,
3300 * switched_from_dl() will ta 2648 * switched_from_dl() will take care to do it at the correct
3301 * (0-lag) time. 2649 * (0-lag) time.
3302 */ 2650 */
3303 err = 0; 2651 err = 0;
3304 } 2652 }
3305 raw_spin_unlock(&dl_b->lock); 2653 raw_spin_unlock(&dl_b->lock);
3306 2654
3307 return err; 2655 return err;
3308 } 2656 }
3309 2657
3310 /* 2658 /*
3311 * This function initializes the sched_dl_ent 2659 * This function initializes the sched_dl_entity of a newly becoming
3312 * SCHED_DEADLINE task. 2660 * SCHED_DEADLINE task.
3313 * 2661 *
3314 * Only the static values are considered here 2662 * Only the static values are considered here, the actual runtime and the
3315 * absolute deadline will be properly calcula 2663 * absolute deadline will be properly calculated when the task is enqueued
3316 * for the first time with its new policy. 2664 * for the first time with its new policy.
3317 */ 2665 */
3318 void __setparam_dl(struct task_struct *p, con 2666 void __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3319 { 2667 {
3320 struct sched_dl_entity *dl_se = &p->d 2668 struct sched_dl_entity *dl_se = &p->dl;
3321 2669
3322 dl_se->dl_runtime = attr->sched_runti 2670 dl_se->dl_runtime = attr->sched_runtime;
3323 dl_se->dl_deadline = attr->sched_dead 2671 dl_se->dl_deadline = attr->sched_deadline;
3324 dl_se->dl_period = attr->sched_period 2672 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3325 dl_se->flags = attr->sched_flags & SC !! 2673 dl_se->flags = attr->sched_flags;
3326 dl_se->dl_bw = to_ratio(dl_se->dl_per 2674 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3327 dl_se->dl_density = to_ratio(dl_se->d 2675 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3328 } 2676 }
3329 2677
3330 void __getparam_dl(struct task_struct *p, str 2678 void __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3331 { 2679 {
3332 struct sched_dl_entity *dl_se = &p->d 2680 struct sched_dl_entity *dl_se = &p->dl;
3333 2681
3334 attr->sched_priority = p->rt_priority 2682 attr->sched_priority = p->rt_priority;
3335 attr->sched_runtime = dl_se->dl_runti 2683 attr->sched_runtime = dl_se->dl_runtime;
3336 attr->sched_deadline = dl_se->dl_dead 2684 attr->sched_deadline = dl_se->dl_deadline;
3337 attr->sched_period = dl_se->dl_period 2685 attr->sched_period = dl_se->dl_period;
3338 attr->sched_flags &= ~SCHED_DL_FLAGS; !! 2686 attr->sched_flags = dl_se->flags;
3339 attr->sched_flags |= dl_se->flags; <<
3340 } 2687 }
3341 2688
3342 /* 2689 /*
>> 2690 * Default limits for DL period; on the top end we guard against small util
>> 2691 * tasks still getting rediculous long effective runtimes, on the bottom end we
>> 2692 * guard against timer DoS.
>> 2693 */
>> 2694 unsigned int sysctl_sched_dl_period_max = 1 << 22; /* ~4 seconds */
>> 2695 unsigned int sysctl_sched_dl_period_min = 100; /* 100 us */
>> 2696
>> 2697 /*
3343 * This function validates the new parameters 2698 * This function validates the new parameters of a -deadline task.
3344 * We ask for the deadline not being zero, an 2699 * We ask for the deadline not being zero, and greater or equal
3345 * than the runtime, as well as the period of 2700 * than the runtime, as well as the period of being zero or
3346 * greater than deadline. Furthermore, we hav 2701 * greater than deadline. Furthermore, we have to be sure that
3347 * user parameters are above the internal res 2702 * user parameters are above the internal resolution of 1us (we
3348 * check sched_runtime only since it is alway 2703 * check sched_runtime only since it is always the smaller one) and
3349 * below 2^63 ns (we have to check both sched 2704 * below 2^63 ns (we have to check both sched_deadline and
3350 * sched_period, as the latter can be zero). 2705 * sched_period, as the latter can be zero).
3351 */ 2706 */
3352 bool __checkparam_dl(const struct sched_attr 2707 bool __checkparam_dl(const struct sched_attr *attr)
3353 { 2708 {
3354 u64 period, max, min; 2709 u64 period, max, min;
3355 2710
3356 /* special dl tasks don't actually us 2711 /* special dl tasks don't actually use any parameter */
3357 if (attr->sched_flags & SCHED_FLAG_SU 2712 if (attr->sched_flags & SCHED_FLAG_SUGOV)
3358 return true; 2713 return true;
3359 2714
3360 /* deadline != 0 */ 2715 /* deadline != 0 */
3361 if (attr->sched_deadline == 0) 2716 if (attr->sched_deadline == 0)
3362 return false; 2717 return false;
3363 2718
3364 /* 2719 /*
3365 * Since we truncate DL_SCALE bits, m 2720 * Since we truncate DL_SCALE bits, make sure we're at least
3366 * that big. 2721 * that big.
3367 */ 2722 */
3368 if (attr->sched_runtime < (1ULL << DL 2723 if (attr->sched_runtime < (1ULL << DL_SCALE))
3369 return false; 2724 return false;
3370 2725
3371 /* 2726 /*
3372 * Since we use the MSB for wrap-arou 2727 * Since we use the MSB for wrap-around and sign issues, make
3373 * sure it's not set (mind that perio 2728 * sure it's not set (mind that period can be equal to zero).
3374 */ 2729 */
3375 if (attr->sched_deadline & (1ULL << 6 2730 if (attr->sched_deadline & (1ULL << 63) ||
3376 attr->sched_period & (1ULL << 63) 2731 attr->sched_period & (1ULL << 63))
3377 return false; 2732 return false;
3378 2733
3379 period = attr->sched_period; 2734 period = attr->sched_period;
3380 if (!period) 2735 if (!period)
3381 period = attr->sched_deadline 2736 period = attr->sched_deadline;
3382 2737
3383 /* runtime <= deadline <= period (if 2738 /* runtime <= deadline <= period (if period != 0) */
3384 if (period < attr->sched_deadline || 2739 if (period < attr->sched_deadline ||
3385 attr->sched_deadline < attr->sche 2740 attr->sched_deadline < attr->sched_runtime)
3386 return false; 2741 return false;
3387 2742
3388 max = (u64)READ_ONCE(sysctl_sched_dl_ 2743 max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
3389 min = (u64)READ_ONCE(sysctl_sched_dl_ 2744 min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
3390 2745
3391 if (period < min || period > max) 2746 if (period < min || period > max)
3392 return false; 2747 return false;
3393 2748
3394 return true; 2749 return true;
3395 } 2750 }
3396 2751
3397 /* 2752 /*
3398 * This function clears the sched_dl_entity s 2753 * This function clears the sched_dl_entity static params.
3399 */ 2754 */
3400 static void __dl_clear_params(struct sched_dl !! 2755 void __dl_clear_params(struct task_struct *p)
3401 { 2756 {
>> 2757 struct sched_dl_entity *dl_se = &p->dl;
>> 2758
3402 dl_se->dl_runtime = 0; 2759 dl_se->dl_runtime = 0;
3403 dl_se->dl_deadline = 0; 2760 dl_se->dl_deadline = 0;
3404 dl_se->dl_period = 0; 2761 dl_se->dl_period = 0;
3405 dl_se->flags = 0; 2762 dl_se->flags = 0;
3406 dl_se->dl_bw = 0; 2763 dl_se->dl_bw = 0;
3407 dl_se->dl_density = 0; 2764 dl_se->dl_density = 0;
3408 2765
>> 2766 dl_se->dl_boosted = 0;
3409 dl_se->dl_throttled = 0; 2767 dl_se->dl_throttled = 0;
3410 dl_se->dl_yielded = 0; 2768 dl_se->dl_yielded = 0;
3411 dl_se->dl_non_contending = 0; 2769 dl_se->dl_non_contending = 0;
3412 dl_se->dl_overrun = 0; 2770 dl_se->dl_overrun = 0;
3413 dl_se->dl_server = 0; <<
3414 <<
3415 #ifdef CONFIG_RT_MUTEXES <<
3416 dl_se->pi_se = dl_ <<
3417 #endif <<
3418 } <<
3419 <<
3420 void init_dl_entity(struct sched_dl_entity *d <<
3421 { <<
3422 RB_CLEAR_NODE(&dl_se->rb_node); <<
3423 init_dl_task_timer(dl_se); <<
3424 init_dl_inactive_task_timer(dl_se); <<
3425 __dl_clear_params(dl_se); <<
3426 } 2771 }
3427 2772
3428 bool dl_param_changed(struct task_struct *p, 2773 bool dl_param_changed(struct task_struct *p, const struct sched_attr *attr)
3429 { 2774 {
3430 struct sched_dl_entity *dl_se = &p->d 2775 struct sched_dl_entity *dl_se = &p->dl;
3431 2776
3432 if (dl_se->dl_runtime != attr->sched_ 2777 if (dl_se->dl_runtime != attr->sched_runtime ||
3433 dl_se->dl_deadline != attr->sched 2778 dl_se->dl_deadline != attr->sched_deadline ||
3434 dl_se->dl_period != attr->sched_p 2779 dl_se->dl_period != attr->sched_period ||
3435 dl_se->flags != (attr->sched_flag !! 2780 dl_se->flags != attr->sched_flags)
3436 return true; 2781 return true;
3437 2782
3438 return false; 2783 return false;
3439 } 2784 }
3440 2785
3441 #ifdef CONFIG_SMP 2786 #ifdef CONFIG_SMP
>> 2787 int dl_task_can_attach(struct task_struct *p, const struct cpumask *cs_cpus_allowed)
>> 2788 {
>> 2789 unsigned long flags, cap;
>> 2790 unsigned int dest_cpu;
>> 2791 struct dl_bw *dl_b;
>> 2792 bool overflow;
>> 2793 int ret;
>> 2794
>> 2795 dest_cpu = cpumask_any_and(cpu_active_mask, cs_cpus_allowed);
>> 2796
>> 2797 rcu_read_lock_sched();
>> 2798 dl_b = dl_bw_of(dest_cpu);
>> 2799 raw_spin_lock_irqsave(&dl_b->lock, flags);
>> 2800 cap = dl_bw_capacity(dest_cpu);
>> 2801 overflow = __dl_overflow(dl_b, cap, 0, p->dl.dl_bw);
>> 2802 if (overflow) {
>> 2803 ret = -EBUSY;
>> 2804 } else {
>> 2805 /*
>> 2806 * We reserve space for this task in the destination
>> 2807 * root_domain, as we can't fail after this point.
>> 2808 * We will free resources in the source root_domain
>> 2809 * later on (see set_cpus_allowed_dl()).
>> 2810 */
>> 2811 int cpus = dl_bw_cpus(dest_cpu);
>> 2812
>> 2813 __dl_add(dl_b, p->dl.dl_bw, cpus);
>> 2814 ret = 0;
>> 2815 }
>> 2816 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
>> 2817 rcu_read_unlock_sched();
>> 2818
>> 2819 return ret;
>> 2820 }
>> 2821
3442 int dl_cpuset_cpumask_can_shrink(const struct 2822 int dl_cpuset_cpumask_can_shrink(const struct cpumask *cur,
3443 const struct 2823 const struct cpumask *trial)
3444 { 2824 {
3445 unsigned long flags, cap; !! 2825 int ret = 1, trial_cpus;
3446 struct dl_bw *cur_dl_b; 2826 struct dl_bw *cur_dl_b;
3447 int ret = 1; !! 2827 unsigned long flags;
3448 2828
3449 rcu_read_lock_sched(); 2829 rcu_read_lock_sched();
3450 cur_dl_b = dl_bw_of(cpumask_any(cur)) 2830 cur_dl_b = dl_bw_of(cpumask_any(cur));
3451 cap = __dl_bw_capacity(trial); !! 2831 trial_cpus = cpumask_weight(trial);
>> 2832
3452 raw_spin_lock_irqsave(&cur_dl_b->lock 2833 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
3453 if (__dl_overflow(cur_dl_b, cap, 0, 0 !! 2834 if (cur_dl_b->bw != -1 &&
>> 2835 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
3454 ret = 0; 2836 ret = 0;
3455 raw_spin_unlock_irqrestore(&cur_dl_b- 2837 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
3456 rcu_read_unlock_sched(); 2838 rcu_read_unlock_sched();
3457 2839
3458 return ret; 2840 return ret;
3459 } 2841 }
3460 2842
3461 enum dl_bw_request { !! 2843 bool dl_cpu_busy(unsigned int cpu)
3462 dl_bw_req_check_overflow = 0, <<
3463 dl_bw_req_alloc, <<
3464 dl_bw_req_free <<
3465 }; <<
3466 <<
3467 static int dl_bw_manage(enum dl_bw_request re <<
3468 { 2844 {
3469 unsigned long flags; !! 2845 unsigned long flags, cap;
3470 struct dl_bw *dl_b; 2846 struct dl_bw *dl_b;
3471 bool overflow = 0; !! 2847 bool overflow;
3472 2848
3473 rcu_read_lock_sched(); 2849 rcu_read_lock_sched();
3474 dl_b = dl_bw_of(cpu); 2850 dl_b = dl_bw_of(cpu);
3475 raw_spin_lock_irqsave(&dl_b->lock, fl 2851 raw_spin_lock_irqsave(&dl_b->lock, flags);
3476 !! 2852 cap = dl_bw_capacity(cpu);
3477 if (req == dl_bw_req_free) { !! 2853 overflow = __dl_overflow(dl_b, cap, 0, 0);
3478 __dl_sub(dl_b, dl_bw, dl_bw_c <<
3479 } else { <<
3480 unsigned long cap = dl_bw_cap <<
3481 <<
3482 overflow = __dl_overflow(dl_b <<
3483 <<
3484 if (req == dl_bw_req_alloc && <<
3485 /* <<
3486 * We reserve space i <<
3487 * root_domain, as we <<
3488 * We will free resou <<
3489 * later on (see set_ <<
3490 */ <<
3491 __dl_add(dl_b, dl_bw, <<
3492 } <<
3493 } <<
3494 <<
3495 raw_spin_unlock_irqrestore(&dl_b->loc 2854 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
3496 rcu_read_unlock_sched(); 2855 rcu_read_unlock_sched();
3497 2856
3498 return overflow ? -EBUSY : 0; !! 2857 return overflow;
3499 } <<
3500 <<
3501 int dl_bw_check_overflow(int cpu) <<
3502 { <<
3503 return dl_bw_manage(dl_bw_req_check_o <<
3504 } <<
3505 <<
3506 int dl_bw_alloc(int cpu, u64 dl_bw) <<
3507 { <<
3508 return dl_bw_manage(dl_bw_req_alloc, <<
3509 } <<
3510 <<
3511 void dl_bw_free(int cpu, u64 dl_bw) <<
3512 { <<
3513 dl_bw_manage(dl_bw_req_free, cpu, dl_ <<
3514 } 2858 }
3515 #endif 2859 #endif
3516 2860
3517 #ifdef CONFIG_SCHED_DEBUG 2861 #ifdef CONFIG_SCHED_DEBUG
3518 void print_dl_stats(struct seq_file *m, int c 2862 void print_dl_stats(struct seq_file *m, int cpu)
3519 { 2863 {
3520 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl) 2864 print_dl_rq(m, cpu, &cpu_rq(cpu)->dl);
3521 } 2865 }
3522 #endif /* CONFIG_SCHED_DEBUG */ 2866 #endif /* CONFIG_SCHED_DEBUG */
3523 2867
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.